[Bug] CUDA extension failures terminate the Python process

[morluto]

Problem

The CUDA extension's CUCHECK and CUDACHECK macros print an error and call
exit(EXIT_FAILURE). Any checked CUDA driver or runtime failure therefore
terminates the Python process from inside the library.

Python caller --> pybind binding --> CUDA error --> exit()
                                              |
                                              `--> process ends

The caller cannot catch the error, attach operation context, clean up its own
state, or coordinate an orderly distributed shutdown.

Reproduction

Pass nvl_multicast_import a valid POSIX file descriptor for /dev/null. The
descriptor is valid at the OS boundary but is not a CUDA-exported shareable
allocation, so the CUDA import call returns an error.

On an H100 NVL with origin/master, the child process printed the old macro's
CUDA error and exited with status 1. The parent regression failed because no
Python exception was catchable:

assert result.returncode == 0
actual returncode: 1
stderr: Failed: CUDA error ... 'invalid device ordinal'

This is a defined error path: it does not fabricate or reinterpret an opaque
CUDA allocation handle.

Expected behavior

• Checked CUDA calls raise a Python-visible RuntimeError through pybind.
• The message identifies the failed expression, symbolic error name, numeric
  code, and CUDA description when available.
• Partially acquired VMM resources are released while the C++ stack unwinds.
• Successful allocation, mapping, and multicast ownership behavior remains
  unchanged.

Python caller --> pybind binding --> CUDA error --> RuntimeError
                                              |
                                              `--> scoped VMM cleanup

Some asynchronous CUDA failures can leave a CUDA context unusable, and the
application may still decide to terminate. That recovery policy should remain
with the caller rather than being enforced by a library macro for every CUDA
error.