> In general, `cuda.core` has a few internal, module-level states that should be guarded by a lock. We haven't done that extensively. _Originally posted by @leofang in https://github.com/NVIDIA/cuda-python/pull/1364#discussion_r2624683212_