chore(tests): stop leaking channels on cleanup failure

[nijeesh-stream]

Summary

Audit: the CI suite creates ~94 channels per run (~50 sync + ~44 async) across fixtures and explicit test creations. Cleanup historically leaked them three ways:

#	Source	Channels / run
1	channel fixture (scope="function", used by 37 sync + 31 async tests)	68
2	fellowship_of_the_ring fixture	2
3	test_channel.py explicit creates (lines 32, 43, 322)	3 sync + 3 async = 6
4	test_draft.py:88 / async test_draft.py:88 (channel2, no cleanup on failure)	1 sync + 1 async = 2
5	test_channel_batch_updater.py (4 tests × ch1+ch2, both suites)	16

Three concrete leak vectors:

1. channel fixture used client.delete_channels([cid], hard_delete=True) — that returns a task_id and the fixture returns immediately. When the task queue is backed up (the same backend slowness that breaks test_delete_channels) the channel is still alive when the next test asserts on a clean app.
2. Every cleanup path was except Exception: pass — failures invisible. CI logs show nothing when delete fails.
3. test_query_drafts (sync + async) created channel2 but only cleaned it up on the happy path — any assertion failure mid-test leaked it forever.

Stale channels drift across runs and eventually break unrelated tests with stale-state assertions (e.g. len(...) == 3 got 15).

Fix

Fix	Effect
A. Swap client.delete_channels(...) → channel.delete(hard=True) in every fixture. Synchronous HTTP DELETE, no task_id race.	Channel gone before fixture returns.
C. Log cleanup exceptions to stderr ([cleanup] channel <cid> failed: <kind>: <msg>).	Failures visible in CI logs, don't fail the passing test.
D. New _purge_stale_test_channels session-scoped, autouse=True fixture queries channels tagged {"test": True, "language": "python"} at session start and end and hard-deletes them. Tag test_query_drafts's channel2 so the sweep catches it too.	Self-healing across runs.

Test plan

• uv run black — clean
• CI: green on the matrix (3.8 / 3.9 / 3.10 / 3.11 / 3.12)
• CI logs: [cleanup] lines only appear when something legitimately failed; on a clean run there should be at most one [cleanup] swept N leaked test channels line at session start (covering whatever the prior run left behind) and possibly one at session end (own-run strays, if any).