fix: await async def setup() coroutine instead of silently dropping it

[michaeldwan]

When a predictor has async def setup(), coglet calls it but doesn't await the coroutine. Setup appears to succeed -- logs say "Setup complete" and "Server ready" -- but none of the setup code actually executes. Every prediction then fails with AttributeError because attributes were never set.

call_method0("setup") returns a coroutine object when setup is async. The ? only checks for Python exceptions, so the coroutine was silently dropped.

predict() and train() don't have this problem because load() calls detect_async() for both and call_method_raw() runs the coroutine with asyncio.run(). That pattern was never applied to setup().

The fix: detect whether setup() is async during load(), capture the return value in setup(), and run it with asyncio.run() when it's a coroutine.

Three integration tests cover async setup with async predict, sync predict, and the weights parameter path.

Fixes #2919

[ask-bonk]

I'm Bonk, and I've done a quick review of your PR.

This PR fixes a critical bug where async def setup() coroutines were silently dropped instead of being awaited, causing predictions to fail with AttributeError because setup attributes were never set.

The fix follows the same pattern used for predict() and train():

1. Detect async setup during load() using detect_async()
2. Store result from setup() call
3. Conditionally run with asyncio.run() when setup_is_async is true

LGTM

github run

[markphelps]

nice

[ask-bonk]

LGTM

github run