feat(kibana): send session id as prompt_cache_key for prompt caching - #122

…patible endpoint (elastic#294736)

## Summary

The inference plugin already supports `sessionId` + `cacheControl` for
EIS prompt caching (elastic#284774), but the Elastic Console / Elastic Ramen
OpenAI-compatible endpoint
(`/internal/elastic_ramen/v1/chat/completions`) didn't forward them, so
external agents (Elastic Ramen, pi, …) never got prompt caching.

This PR adds that wiring:

- **Session id** is resolved from (first match wins):
1. `prompt_cache_key` body field (standard OpenAI field; used by Elastic
Ramen via `@ai-sdk/openai-compatible`)
2. `x-session-id` header (OpenRouter convention; pi sends this with
`sessionAffinityFormat: "openrouter"`)
  3. `x-session-affinity` header
- When a session id is present, `cacheControl: { type: 'ephemeral', ttl:
'5m' }` is set, matching the Agent Builder main loop.
`prompt_cache_retention: "24h"` maps to the longest EIS TTL (`1h`).
- Usage now includes `prompt_tokens_details.cached_tokens` whenever
inference reports cached tokens, so clients like ai-sdk and pi can show
cache reads.
- Session ids are bounded to 256 chars (validated for the body field;
oversized header values are ignored).
- **Request body limit raised to 20MB** for `/chat/completions` (from
the 1MB `server.maxPayload` default). Long agent conversations with
large tool results or base64 images quickly go over 1MB, and every turn
resends the full history.
- **Message limits raised**: up to 10,000 messages per request (was
1,000) and 2,000 content parts per message (was 100). Each tool call
adds an assistant message and a tool message, so long agent sessions hit
1,000 well before the context window fills. The total payload is still
capped by the 20MB body limit.
- README documents prompt caching and shows a verified pi `models.json`
setup.

Companion Elastic Ramen PR: elastic/elastic-ramen#122

### Verification (local stack with EIS)

Checked through Phoenix traces (`elastic.cache_control.*` span
attributes):
- `prompt_cache_key` in the body → `session_id` + `ttl: 5m` ✅
- `x-session-id` header → `session_id` + `ttl: 5m` ✅
- pi (`openai-completions` + compat config from README) → every turn of
a pi session carries the pi session id ✅. With `PI_CACHE_RETENTION=long`
→ `ttl: 1h` ✅

Body limit: checked live. A ~5MB request returns `200`; a ~21MB request
returns `413 Payload content length greater than maximum allowed:
20971520`.

**Prompt cache hits (Claude Sonnet 5 on EIS):** I sent the same
~422k-token request 3 times with one `prompt_cache_key`:

| Request | Time | `cached_tokens` |
|---|---|---|
| #1 | 8.3s | – (cache written) |
| #2 | 2.1s | 422,347 / 422,349 |
| #3 | 2.0s | 422,347 / 422,349 |

So the session id reaches EIS, EIS caches the prompt, and the new
`prompt_tokens_details.cached_tokens` field reports the hit. Note:
Claude Sonnet 4.5 on EIS doesn't report cached tokens, so no hits show
up for that model.

**Large inputs** (one-word answer hidden in the middle of the input,
sent through the route):

| Model | Prompt tokens | Result |
|---|---|---|
| Claude Sonnet 5 | 422k / 814k | ✅ / ✅ |
| Gemini 2.5 Pro | 388k / 752k | ✅ / ✅ |
| GPT-5.4 | 291k | ✅ |
| Claude Sonnet 4.5 | 124k | ✅ (EIS context window looks like 200k) |

**Long conversations:** a request with 5,001 messages returns `200`.
That was over the old 1,000-message limit.

### Checklist

- [x] [Unit or functional
tests](https://www.elastic.co/guide/en/kibana/master/development-tests.html)
were updated or added to match the most common scenarios
- [x]
[Documentation](https://www.elastic.co/guide/en/kibana/master/development-documentation.html)
was added for features that require explanation or tutorials

## Release Notes

N/A (experimental feature)