This PR started as a small PR to make it easy to create thread local streams in Python. But I think we had the teardown logic quite wrong but it didn't show up until running the server which does the computation in a background dedicated thread.

Basically, thread local objects are torn down when the thread exits which is likely to be before the interpreter exits. In that case the destructors of C++ objects will run and try to clean up Python objects without holding the GIL. So the fix is the standard hold the GIL and clean up in destructor situation.

Regarding the main thread though, the interpreter will probably be torn down before the thread local objects. So when the destructors run there will be no python and no GIL and segfault. So we register exit handlers that clean up and the C++ objects check if there is cleanup to do or do nothing.

The above means that correct usage of MLX requires either manual clean up of some kind, or all threads to join the main thread before exiting.