Clear CUDA errors once reported and raise eval errors in buffer protocol

[aleroot]

Investigating the CI build failure I realized that : a CUDA error that MLX has already reported stays set on the calling thread, and Event::wait() re reads it with cudaPeekAtLastError().
After one failed allocation every later eval on that thread throws cudaPeekAtLastError() failed: out of memory, even though nothing is wrong. When the read comes through the buffer protocol the exception escapes the extern "C" getbuffer slot and the process aborts with exit 134.
This is what took down the Linux (cuda-*, x86_64) jobs after an OOM in test_fast_sdpa. see: #4522

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