Do not throw from CUDA destructors and avoid implicit default streams in eval/compile

[aleroot]

As described in #4506 a termination crash occurs when a detached worker thread outlives the CUDA runtime and attempts to clean up CUDA resources, triggering a std::terminate.

To resolve this I have made some changes so that the worker class is restructured therefore the detached thread only captures a new, inert State struct containing standard threading primitives. Additionally the CUDA resources remain in the Worker, which is now managed by the CommandEncoder as a unique_ptr to ensure deterministic destruction before the runtime unloads. The destructor ~CudaHandle has been modified to safely ignore errors during shutdown, and eval_impl now selects streams lazily to prevent accidental worker instantiation.

Fixes #4506

• ☑️ I understand it is strictly prohibited to use AI to write PR description
• AI usage disclosure:

[zcbenz]

I think there is no need to change Worker, the crash in #4506 was caused by CommandEncoder getting destroyed after CUDA runtime shutdown and this change is not going to prevent that.

[aleroot]

I think there is no need to change Worker, the crash in #4506 was caused by CommandEncoder getting destroyed after CUDA runtime shutdown and this change is not going to prevent that.

@zcbenz you are right on the Worker, so I've dropped the Worker changes. For the record, the #4506 trace unwinds from the detached thread's ~_State_impl into ~Worker, so ~CommandEncoder had already synchronized fine. The crash was the late cudaEventDestroy throwing, and non-throwing destructors cover that.

I think we are good now and hopefully the fixes can then unlock the the new release on mlx-swift. @zcbenz thanks for the support, appreciated.