feat(tsdb): make ES95 the default TSDB codec - #152861

* feat(tsdb): make ES95 the default doc values codec

Remove the `es95_codec` feature flag and turn the ES95 TSDB doc values
codec on by default for time series indices on stateful nodes. The
`index.time_series.es95_codec.enabled` setting is kept so the codec can
be disabled at index creation.

The default is gated on a new index version,
`TIME_SERIES_ES95_CODEC_DEFAULT_STATEFUL`, so only indices created after
this change use ES95 while existing indices keep ES819.

Serverless is held back: the ES95 codec is disabled on stateless nodes.
Because `stateless.enabled` is a node-scoped setting, this decision is
made in the `IndexSettings` constructor, where the node settings are
available. It will be enabled later through a dedicated
`TIME_SERIES_ES95_CODEC_DEFAULT_STATELESS` index version.

* docs: add changelog for ES95 default codec on stateful

* docs: drop stateful mention from ES95 changelog

* feat(tsdb): add cluster setting to gate the ES95 codec

Add `cluster.time_series.es95_codec.enabled`, a dynamic cluster setting
that disables the ES95 TSDB doc values codec cluster-wide. When set to
`false`, newly created time series indices force
`index.time_series.es95_codec.enabled` to `false`, so the codec is
disabled on rollover while existing indices keep their codec.

The value is applied at index creation through a builtin
`IndexSettingProvider` that caches it from `ClusterSettings`, so a
runtime change is picked up by the next index creation without a
restart.

* refactor(tsdb): drop stateful from ES95 constant

Rename `TIME_SERIES_ES95_CODEC_DEFAULT_STATEFUL` to
`TIME_SERIES_ES95_CODEC_DEFAULT` and drop `stateful` from the related
test names. The version keys the codec default for all node types; the
stateless holdback is handled separately in the `IndexSettings`
constructor.

* test(tsdb): cover cluster kill switch for standalone index

Add an integration test that creates a standalone time series index
through the create index API, with `cluster.time_series.es95_codec.enabled`
set to `false`, and asserts the codec is forced off. The existing tests
only exercised the data stream path.

* Update docs/changelog/152861.yaml

* docs: set ES95 changelog type to feature

The `release highlight` label bot set `type: non-issue`, which is not in
the changelog schema enum and failed `validateChangelogs`. Use `feature`,
a valid type; combined with the highlight the change appears in both the
features list and the release highlights.

* docs: outline ES95 compression gains in highlight

elastic#152861 retired the es95_codec feature flag and made ES95 the default
TSDB doc values codec, so the lifecycle IT no longer needs to enable
the flag on its test cluster.

* feat(tsdb): default ES95 codec on serverless

Enable the ES95 TSDB doc values codec by default on stateless nodes,
where #152861 held it back. Drop the blanket stateless holdback in the
`IndexSettings` constructor and gate stateless enablement on a new index
version, `TIME_SERIES_ES95_CODEC_DEFAULT_STATELESS`, so existing
serverless indices keep ES819 and only indices created on or after it
use ES95.

Stateful behavior is unchanged: its default still gates on
`TIME_SERIES_ES95_CODEC_DEFAULT`. The per-index opt-out and the cluster
wide kill switch keep working on serverless, since an explicit `false`
short-circuits the codec regardless of node type.

* refactor(tsdb): drop stateless ES95 version gate

`TIME_SERIES_ES95_CODEC_DEFAULT` has not reached serverless yet, so
gate serverless on it directly and drop the dedicated
`TIME_SERIES_ES95_CODEC_DEFAULT_STATELESS`.

* fix(tsdb): make ES95 codec setting dynamic and classify in CCR

Drop `Property.Final` and add `Property.Dynamic` to
`TIME_SERIES_ES95_CODEC_ENABLED_SETTING` so the setting can be flipped on a
live backing index. This enables the in place migration and rollback paths
exercised by the data stream lifecycle tests, where a single write backing
index switches codec via force merge after a setting update.

Once the setting is dynamic, `TransportResumeFollowActionTests` requires
that it be classified as replicated or non replicated. The codec opt in is
a per index choice and must not follow from leader to follower, so add it
to `NON_REPLICATED_SETTINGS`.

* test(tsdb): add ES95 data stream lifecycle tests

Adds seven REST tests for the ES95 TSDB codec inside TSDB data streams.
Each scenario uses a manual rollover or a dynamic setting flip to exercise
the codec switch path that production users follow when migrating to ES95
in place.

Covered scenarios:
* baseline to ES95 rollover, ES95 to baseline rollover, four way codec
  alternation across rollovers
* in place flip to ES95 on the write backing index followed by force merge
* in place flip back to baseline on the write backing index followed by
  force merge
* alias level queries across mixed codec backing indices
* random access doc values across mixed codec backing indices

Per segment format correctness is covered by `ES95VsES819DocValuesDuelTests`.
Cluster restart and rolling upgrade safety are covered by the matching
javaRestTest suites on their own PRs.

* revert(tsdb): restore Property.Final on ES95 codec setting

Roll back `TIME_SERIES_ES95_CODEC_ENABLED_SETTING` to its original
`Property.IndexScope` plus `Property.Final` after review feedback
preferring the historical convention of immutable codec settings.
The setting is fixed at index creation and cannot be updated on a
live index.

Drop the CCR `NON_REPLICATED_SETTINGS` entry, only required while the
setting was dynamic, and the two tests that flip the setting on a
live write backing index: `testInPlaceFlipToES95OnWriteBackingIndex`
and `testInPlaceFlipToBaselineOnWriteBackingIndex`. The
`updateIndexSetting` and `forceMerge` helpers have no remaining
callers and are removed with them.

* test(tsdb): wire ES95 codec feature flag in data stream lifecycle cluster

* test(tsdb): drop es95_codec flag from lifecycle IT

#152861 retired the es95_codec feature flag and made ES95 the default
TSDB doc values codec, so the lifecycle IT no longer needs to enable
the flag on its test cluster.

* test(tsdb): make ES95 lifecycle IT stateless-safe

The serverless CI re-runs this IT against a stateless cluster, where a
new backing index's search shards are loaded from the object store
asynchronously. The count read at the start of bulkIndex raced that
loading and failed with a 503 no_shard_available_action_exception.

Wait for the write index with ensureGreen before that read.

* test(tsdb): drop replica pin in ES95 lifecycle IT

The serverless CI re-runs this IT against a stateless cluster, where
reads are served by a separate search shard. Pinning
index.number_of_replicas to 0 left the search node without a shard, so
the count and search calls failed with a 503
no_shard_available_action_exception.

Drop the pin so stateless can provision the search shard.

* test(tsdb): wait for a searchable shard, not green, in ES95 IT

Dropping the replica pin fixed serverless but broke the single node
stateful run: the default replica stays unassigned, the cluster sits at
yellow, and ensureGreen times out on every backing index.

Replace ensureGreen with a probe that retries a search until the shard
can serve it. This fits both cluster shapes: stateful serves from the
primary at once, and stateless retries until its asynchronously
provisioned search shard is ready. Cap the wait at 60s so a stuck shard
fails the test instead of blocking.