Python(fix): sift client lifecycle improvements

[alexluck-sift]

Changes

• GrpcClient.close_sync() / close() now clear _channels_async and _stubs_async_map after teardown so the channel objects are released promptly instead of lingering past interpreter exit.
• close_sync() is now idempotent (guards against double-close), skips work when the loop is already stopped, and uses a .result(timeout=...) so an unresponsive channel can't block forever at exit.
• Added TestGrpcClientClose covering the ref clearing and the idempotent second-call paths.

Why

The pytest plugin (and any long-lived sync caller) was leaving grpc_wait_for_shutdown_with_timeout() timed out at process exit, plus a silent ~2s stall while gRPC waited.

The gRPC C-core releases a channel's resources only when the Python channel object is destroyed, not when close() returns. Under pytest the session-scoped client kept channels alive in its maps until interpreter finalization, so the C-core's own exit-time shutdown raced live channels and timed out. Clearing the maps inside close() drops the refs, refcounting destroys the channels, and the C-core's later C-level atexit then finds nothing to wait on.

The idempotency guard fixes a separate latent hang: closing explicitly (or via the context manager) and then having atexit close again submitted a coroutine to the already-stopped loop, where .result() blocked forever.

[github-actions]

Python docs preview: https://sift-stack.github.io/sift/python/pr-595/

Deployed from d945180. The link may take up to a minute to become live as GitHub Pages propagates.