fix: async setup() and predict() run on different event loops

[michaeldwan]

Async predictors that create event-loop-bound resources in setup() -- httpx.AsyncClient, aiohttp.ClientSession, asyncio.Queue, anything that binds to the running loop -- crash on 0.17 because async setup() ran via asyncio.run() (ephemeral loop, destroyed after setup finishes) while async predict() runs on a separate persistent shared loop via run_coroutine_threadsafe.

The fix submits async setup coroutines to the same shared event loop that predict uses, so both phases share one loop. Sync setup is unchanged. The non-worker path (cog predict CLI) keeps using asyncio.run() since it doesn't persist resources across calls.

# before
same_loop=False item=from-setup

# after
same_loop=True item=from-setup

Fixes #2926

[ask-bonk]

LGTM - the fix correctly ensures async setup() and predict() share the same event loop by passing the shared loop to setup and using run_coroutine_threadsafe. The test verifies event-loop-bound resources created in setup work correctly in predict.

[ask-bonk]

I reviewed PR #2927 which fixes a bug where async setup() and predict() ran on different event loops. I read all three changed files (predictor.rs, worker_bridge.rs, test_coglet.py), verified the fix correctly passes the shared event loop to setup for async coroutines, and confirmed the test validates event-loop-bound resources work correctly across both phases. I initially flagged a potential cancellation handling issue but realized setup can't be cancelled via the slot mechanism, then submitted an LGTM review via gh api.

github run