feat: Add BufferedFirestoreSessionService (batched, buffered Firestore session backend)

## What this is

ADK already ships a Firestore session service (`google.adk.integrations.firestore.FirestoreSessionService`). It works well, but it does one Firestore transaction per event — so if 10 messages fire 10 events, that's 10 separate writes, each touching the session document to bump its revision.

This adds a `BufferedFirestoreSessionService` to the community package that holds events in memory and flushes them all in a **single transaction** when the buffer fills up (or on a timer, or on shutdown). The session document gets updated once instead of 10 times, and the state docs get merged once.

## When to use it

You want this if:
- You're on a budget and Firestore costs matter (fewer writes = lower bill)
- Your service runs stably and you can afford to lose the last few seconds of events if it crashes

You want the builtin instead if:
- You need every event durable the moment it happens
- You're running in an environment where the process can be killed without warning

Both services use the same Firestore layout, so they're compatible with each other's data.

## What's in the PR

- `BufferedFirestoreSessionService` in `google.adk_community.sessions`
- Configurable buffer size, flush interval, retry behaviour
- `durable_mode=True` flag to fall back to per-event writes (same as builtin)
- `flat_layout=True` option for a flat `root_collection/{session_id}` layout instead of the default ADK nested path
- All collection names configurable if your Firestore uses a custom structure
- 20 unit tests, no external dependencies
- New `[firestore]` optional extra (`google-cloud-firestore>=2.11,<3`, matching ADK's own constraint)

PR: #160

Review PR #160 alongside the proposed google.adk_community.sessions entry point. Check the 20 described unit tests and the Firestore optional extra, then verify buffered, timer, shutdown, durable-mode, layout, collection, and retry behavior match the issue before considering the work done.

Written by the indexing model from the issue text.