Fix scheduler exception propagation to Python

[Anri-Lombard]

Summary

• Fixed exception propagation from CPU scheduler worker threads to Python
• Added std::exception_ptr storage in StreamThread to capture exceptions thrown during task execution
• Exceptions are now re-thrown during synchronize() so they propagate as RuntimeError to Python

Problem

When mx.linalg.solve is called with a singular matrix, the LU factorization throws a std::runtime_error inside the scheduler worker thread. Previously, this exception was not caught, causing the process to crash with SIGABRT instead of raising a catchable Python exception.

Test plan

• Added regression test in test_linalg.py::test_solve that verifies singular matrix raises RuntimeError
• All existing linalg tests pass (16 tests, 181 subtests)
• Manual verification:

import mlx.core as mx
result = mx.linalg.solve(mx.ones((2, 2)), mx.ones((2,)), stream=mx.cpu)
result.tolist()  # Now raises RuntimeError instead of crashing

Fixes #2888

[awni]

The challenge with this is not so much propagating the exception but making sure the MLX is in a sane state when the exception is raised. Otherwise we don't want to give the client to the opportunity to catch the exception because they may assume it's safe to continue.

Also I think catching the exception only on synchronization is sub-optimal. We should check for exceptions more frequently (like maybe after a wait or something like that.

I'm going to close this for now as I don't think it's at the point that we can consider it. But if you want to send a new PR with the above feedback in mind feel free to do so.