v1.2.6: one bugfix, two behavior improvements

The bug fix is related to preferred read replicas when the read replica
is not a leader of any partition the client is consuming. The behavior
changes are related to metadata refreshes and fetch request partition
ordering.

The details of these commits are a bit involved; see the changelog for
more details or the relevant commits.

v1.2.5: behavior improvements (no hard bug fixes)

A bug fix is in this release, but no broker today validates the behavior
of this bug, meaning nothing can run into it (see KAFKA-13464). This
also handles more errors when transactionally committing so that the
client avoids a fatal state more often. Lastly, tests are easier to run
for anybody.

v1.2.4: fix forgotten topics in fetch sessions, and allow disabling

This patch release fixes a bug with fetch sessions when a topic or
partition would no longer be fetched. There is also a minor bug fix when
using empty layouts for the RecordReader (this is no longer allowed), as
well as some doc improvements.

v1.2.3: fixes not supporting large broker node IDs for admin requests

This patch release fixes a bug that prevented support for metadata
requests against brokers with large node IDs. There are a few doc
improvements, minor beneficial behavior changes, and two tiny new
functions included as well.

v1.2.2: fixes a cooperative consuming bug on repeated fast rebalances

This patch release fixes a bug that could result in partitions not being
consumed during repeated rebalances while cooperative consuming. See the
changelog or the commit for more details.

v1.2.1: fixes a panic after leaving a group

This patch release fixes a panic that can occur during the following
sequence of events:

1) a LeaveGroup occurs
2) the cluster moves a partition from one broker to another
3) a metadata refresh occurs and sees the partition has moved

v1.2.0: record formatting & reading, & minor improvements

This feature release officially tags the new RecordReader and
RecordFormatter types & functions. As well, this contains minor
improvements and behavior changes.

There are no bugfixes in this release.

v1.1.4: make errProducerIDLoadFail retriable

We now avoid failing produced records on retriable init producer id
errors.

v1.1.3: make BrokerNotAvailable retriable, dial errors not retriable

This patch release allows more success while redpanda returns
broker_not_available in some instances, and allows faster fails during
some network issues.

release v1.1.2: patch message set v1 LogAppendTime processing