Fix intermittent crashes at thread and interpreter teardown

[dhiltgen]

Proposed changes

MLX sometimes dies at thread or process teardown with "terminate called without an active exception" or SIGSEGV, on Linux/Windows. The general pattern: a destructor reachable from __call_tls_dtors() must not acquire the GIL, nor throw, nor drop a Python reference.

Acquiring the GIL is fatal because take_gil() answers a request from a non-finalizing thread, while another thread finalizes, by calling PyThread_exit_thread(); the forced unwind then crosses the noexcept destructor frame.

Throwing is equally fatal: CUDA destructors ran CHECK_*_ERROR on teardown calls that legitimately fail -- every Windows run with two threads died.

Adds python/tests/test_teardown.py. The failure kills the process, so the cases run in subprocesses and repeat; both fail without this change.

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)

[zcbenz]

Our current recommended pattern is to let users call clear_streams at the end of threads and the process, so they can do cleanup before cuda and python shutdown, which is something that we could not do on the library side. Have you tried that? We should have a multi-threading doc for all these APIs though.

mlx/tests/scheduler_tests.cpp

Lines 89 to 101 in e78d894

	TEST_CASE("test new stream in threads") {
	std::vector<std::thread> threads;
	for (int i = 0; i < 1; ++i) {
	threads.emplace_back([]() {
	auto s = new_stream(default_device());
	eval(arange(10, s));
	clear_streams();
	});
	}
	for (auto& t : threads) {
	t.join();
	}
	}

[dhiltgen]

I can close this PR, these are just crashes that I was noticing while working on other improvements.