Flush buffered events before the async tracking client's worker exits

[AmirF194]

Fixes #1214.

Changes

BasicAsynchronousHamiltonClient (ui/sdk/src/hamilton_sdk/api/clients.py) had no stop()
and registered no atexit hook, unlike its synchronous sibling. worker()'s only exit path
waits for a None sentinel on data_queue, and nothing ever put one there for the async
client, so a buffered batch sat until the periodic flush timer happened to catch it, or was
lost outright if the process (or the event loop that owned the worker task) exited first.

• Added async def stop() on BasicAsynchronousHamiltonClient: sends the sentinel and awaits
  the worker task so it flushes and exits. This is not wired to atexit, deliberately: the
  worker is a task on the caller's event loop, and by the time atexit callbacks run at
  interpreter shutdown that loop may already be closed, so callers must await stop()
  themselves before the loop stops running (mirroring the sync client isn't possible here for
  that reason; happy to add an atexit-scheduled best-effort variant too if maintainers want
  one, but that has its own shutdown-ordering edge cases worth discussing first).
• Added the matching async def stop() on AsyncHamiltonTracker (adapters.py), delegating
  to the client.
• Fixed a bug in worker()'s sentinel handling that this exercises for the first time: it
  appended the received item to batch before checking whether it was the None sentinel, so
  the first real stop() call flushed a batch containing None and raised. Reordered to check
  first, matching the synchronous client's worker.

How I tested this

Ran in a clean python:3.13-slim and python:3.10-slim Docker container (the SDK's CI matrix
runs 3.10-3.14; I could not run the 3.11/3.12/3.14 legs locally but the change has no
version-specific code paths):

• New regression tests in tests/test_clients.py and tests/test_adapters.py: fail with
  AttributeError on unmodified main (confirmed via git stash), pass on this branch.
  test_async_client_stop_flushes_buffered_item_and_exits_worker enqueues an item, calls
  stop() before the flush interval elapses, and asserts the mocked backend request actually
  fired and the queue drained; it caught the sentinel-ordering bug above on first run.
• Full pytest tests/ (minus the ray/pyspark-gated tests, which are pytest.importorskip
  guarded and unrelated to this change): 140 passed, 5 skipped, no failures.
• ruff check and ruff format --diff on the four changed files: clean. This repo's CI does
  not run a lint/static-checks job for ui/sdk-only changes (checked gh pr checks on a
  recently merged ui/sdk-only PR, fix(ui/sdk): fix failing CI on master due to formatting issues #1626), so this was a courtesy check, not a required gate.

Not verified: behavior under a real ASGI/FastAPI shutdown sequence (e.g. the app in
examples/async/fastapi_example.py, which doesn't call the new stop() either); the fix is
scoped to the tracker itself, not to updating that example.