rust: suppress wake notifications while destroying async tasks

[lthibault]

EVENT_CANCEL skips normal wake-state normalization. If the cancelled task remains SLEEP_STATE_SLEEPING, Rust destructors that wake registered waiters attempt normal notifications after the wake-stream read has been cancelled.

Set SLEEP_STATE_WOKEN at the start of TaskState::drop, before retiring the read and destroying futures. This suppresses destruction-time notifications without scheduling another poll.

tests/runtime/rust-cancel-wake now uses one destructor sender and one Rust oneshot waiter. The runner cancels a sleeping task through the component-model cancellation path, then checks that destruction completed and neither future was polled again. The redundant native test was removed.

Validation:

• The focused component regression passes with the fix, fails at the wake-stream write when only the state store is removed, and passes after restoration. Verified on Wasmtime 48.0.1 and 49.0.2.
• cargo test --locked -p wit-bindgen: default, inter-task-wakeup, async-spawn, combined, and all features; 13 unit tests and 15 doctests per configuration.
• The standard Rust/C runtime suite passes 244 fixture builds and 173 composed executions on both Wasmtime versions.
• Relevant crate tests, workspace Clippy, and formatting pass.

[alexcrichton]

Thanks! Would it be possible to simplify the runtime test to mostly just trying to test the bug being fixed here? As-is it seems pretty unwieldy and I'm having a tough time understanding it

[lthibault]

@alexcrichton, thank you for the prompt reply! I've made the requested changes.

The runtime fixture now has one destructor sender and one oneshot waiter, addressing your review. I removed the sibling task, escaped/repeated wakes, extra gate, and wake graph. One checkpoint remains to let the task sleep before cancellation.

Dropping the import still delivers EVENT_CANCEL. The test checks that destruction completes and neither future is polled again.

[alexcrichton]

Thanks! Reading over this test now it looks like there's a more subtle bug here -- that the spawn_local'd task does not complete looks like a bug, that should still execute tho completion even if the original main future is cancelled/dropped. I'll try to take a closer look at this on Monday

[alexcrichton]

Ok there's definitely a bug, but I think it's unrelated to this, so I'll merge this first.

[lthibault]

Thanks for the review! Are you able to share any details about the bug so that I can keep an eye out for a fix and try to avoid getting tripped up in the meantime?

Cheers!

[alexcrichton]

Oh sure yeah, the fix is #1743