Add register_stream for cross-thread GPU stream usage

[TheTom]

Summary

After #3348 made CommandEncoder thread-local, evaluating arrays on a thread that did not create the stream fails because the encoder does not exist on the new thread. ThreadLocalStream (#3355) solves this for Python by creating per-thread streams, but there is no way to share a single stream across threads.

register_stream(Stream) registers an existing stream's GPU command encoder on the calling thread, allowing eval() to succeed. It is a no-op if already registered, and is safe to call from any thread including the creating thread.

This is needed for Swift concurrency where the runtime may hop tasks between threads, and for Python threading.Thread workflows that share a stream.

Related: #3078

Changes

• mlx/stream.h / mlx/stream.cpp: add register_stream(Stream)
• python/src/stream.cpp: nanobind binding
• python/tests/test_threads.py: cross-thread eval test, idempotency test

Test plan

• python/tests/test_threads.py passes (3/3)
• C++ library builds clean
• CI