cuda.core: should sticky CUDA errors derive from BaseException?

Leggi docs/source/error_handling.rst e il thread di revisione #2759 per conoscere l'attuale policy sticky-error, quindi esamina _check_driver_error, _check_runtime_error e i gestori interni delle eccezioni indicati. Risolvi la decisione sulla gerarchia delle eccezioni e sullo sticky-code prima di occuparti dei test di subprocess, delle esportazioni dell'API, degli stub, della documentazione e delle note di rilascio.

Question

Should cuda.core raise sticky CUDA errors (device faults after which every further CUDA call returns the same error) as a dedicated exception type that derives from BaseException rather than Exception?

This came up in the review of #2759 (see the thread on docs/source/error_handling.rst, "Sticky errors"). Today all driver and runtime errors are raised as cuda.core.CUDAError, which derives from Exception. #2759 documents that sticky errors are raised like any other error and that the application decides how to shut down.

Motivation

After a sticky fault the CUDA documentation calls for the process to be terminated and relaunched; there is no legitimate recovery. A BaseException subclass is not caught by except Exception:, so naive "if anything goes wrong, retry" code would not swallow the fault and keep issuing CUDA work that is guaranteed to fail. It remains catchable for code that explicitly wants to (except BaseException, or the class itself). KeyboardInterrupt and SystemExit are the stdlib precedent for exit-like conditions.

What it would change

• A new public exception type beside CUDAError. It cannot derive from CUDAError, because that would make it an Exception. Consequently except CUDAError: stops catching sticky faults, which is a visible behavior change for existing users.
• _check_driver_error / _check_runtime_error branch on a frozen set of codes. Derived from the shipped error explanations ("any further CUDA work will return the same error"), the driver set is CUDA_ERROR_CONTAINED (226), ILLEGAL_ADDRESS (700), LAUNCH_TIMEOUT (702), HARDWARE_STACK_ERROR (714), ILLEGAL_INSTRUCTION (715), MISALIGNED_ADDRESS (716), INVALID_ADDRESS_SPACE (717), INVALID_PC (718), LAUNCH_FAILED (719), TENSOR_MEMORY_LEAK (721), MPS_CLIENT_TERMINATED (810), EXTERNAL_DEVICE (911). The runtime table has the same codes. CUDA_ERROR_ASSERT is documented as context-fatal, not process-fatal, and would be excluded. A test can cross-check the set against the explanation text.
• Internal except Exception: sites need review, since a BaseException skips them: the Buffer deallocation callback, the free-then-reraise blocks in _memoryview.pyx, the virtual-memory rollback closure, the capture-tail chaining in GraphBuilder. Bare except: sites are unaffected.
• Tests for a real sticky fault must run in a subprocess, because the fault poisons the test process.
• Docs (error_handling.rst, api.rst), __all__, stubs, release notes.

Considerations

• PEP 352 says user-defined exceptions should derive from Exception; direct BaseException subclasses are reserved for conditions that should not be caught by default. A process-fatal device fault arguably qualifies, but this is an unusual stance for a library.
• No ecosystem precedent: CuPy, PyTorch and Numba raise Exception subclasses (RuntimeError or their own) for the same driver codes. In mixed stacks the other library's exception type is what a user sees first anyway.
• pytest, concurrent.futures and threading.excepthook treat a custom BaseException as an ordinary failure. asyncio propagates it out of the event loop, cancelling other tasks, which is the intended "stop" but abrupt.

Alternative

Keep sticky errors as CUDAError (so existing handlers keep working) but raise a dedicated subclass, or set a sticky attribute, so code can recognize them without string matching. This gives the machine-readable signal without the except Exception speed bump.

Decision requested

1. Do we want the BaseException speed bump at the cost of except CUDAError: no longer catching sticky faults?
2. If not, is a CUDAError subclass (or attribute) for sticky faults worth adding?

Refs: #2758 (error-handling policy RFC), #2759 (implementation and review thread).