Fix silent Metal OOM error by sticky stream

[jasonge27]

Proposed changes

Addressing issue #3979, where the MLX drops OOM errors silently in an edge case.

import mlx.core as mx, numpy as np

a = mx.full((3100, 1048576), 3.0)
b = mx.full((3100, 1048576), 4.0)
mx.eval(a, b)
c = a + b
mx.eval(c)
print(float(np.array(c[:1, :1])[0, 0]))   # returns 0.0 (M4 Pro, 48GB)

What goes wrong today:

1. Mid-eval, MLX hits needs_commit() and submits a buffer with no signal events (eval.cpp around the mid-eval commit).
2. That buffer fails → error is parked on the encoder.
3. Nothing is poisoned, because poison only happens for events in signal_events.
4. A later healthy buffer signals the eval synchronizer → mx.eval returns “success.”
5. Output was never written → we get 0.0.

Later, get_command_encoder() / error_.reset() discards the error. In the end, GPU failed, user got silence + wrong data.. I'm also linking a cpp gist to reproduce the error (without having to install python library).

Changes:
Handling error in lazy-eval / async systems is hard. CUDA uses sticky last-error and PyTorch surfaces async errors at the next sync. I think following industry best practices, we should park the error per stream rather than per encoder, and report it from synchronization points rather than from events:

• StreamError keyed by stream index, so it outlives the encode cycle. get_command_encoder() no longer resets it.
• commit() captures the holder by shared_ptr instead of &error_, which also removes a reference that could outlive the encoder.
• Event::wait checks the stream after waiting; EventImpl::wait no longer throws, leaving the event error to propagate across streams via wait_events as before.
• array::wait stops taking theis_available() shortcut, which detached the event as soon as the GPU signaled and skipped the check.
• The cpu-stream branch of Event::wait(Stream) reports nothing. It runs on a scheduler worker with no handler above it, so throwing there terminated the process rather than reaching the user.

With our changes, now the test script above would throw error [METAL] Command buffer execution failed: Insufficient Memory (00000008:kIOGPUCommandBufferCallbackErrorOutOfMemory).

Testing
I added four cases intotests/gpu_tests.cpp. It injects the error into the completion handler, so that we can reproduce the real one needs more memory than a test machine can assume. Reverting either source change fails the matching test. Full suite passes (267 cases).

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

cc @zcbenz

[zcbenz]

Thanks for the PR, while I think your solution works, we are in the middle of refactoring the event handling also for the cpu side (#3742) and I do need to have a minimal fix that would not block too much things, so I'm closing this in favor of #4174.