Handle Windows TCC/WDDM mode more robustly

[leofang]

This issue tracks an internal discussion with QA. This simple snippet shows why using cuda.core today on Windows might fail, depending on if it's TCC or WDDM mode:

>>> from cuda import cuda, cudart
>>> print(cudart.cudaGetDevice())
(<cudaError_t.cudaSuccess: 0>, 0)
>>> print(cuda.cuDeviceGetMemPool(0))
(<CUresult.CUDA_ERROR_NOT_SUPPORTED: 801>, <CUmemoryPool 0x0>)

cuda.core currently assumes the stream-ordered memory allocator is available. However, CUDA on Windows is a bit more complicated than on Linux, since there are two operation modes:

• In the WDDM mode (which is the case during cuda.core development), things should work just fine.
• In the TCC mode (as reported by QA), this is unsupported

We need some treatments to make it usable on TCC.

[leofang]

xref: https://forums.developer.nvidia.com/t/cudaerrornotsupported-during-cudamallocasync-on-windows-10-based-azure-vm-with-tesla-t4-gpu/251405/2

[jrhemstad]

Can't we just say the pool is not available on Windows in TCC mode? I don't think we need to go above and beyond to support something the driver doesn't support.

[leofang]

It is not appropriate because CUDA does support Windows TCC mode, just not the mempool. Right now cuda.core is not functional at all only because I forgot (😞) mempools are not there, but we can easily provide a fallback path to make it work (by wrapping cudaMalloc/cudaFree as suggested in #208).

@jrhemstad I suggest us to take this seriously if we want CUDA Mode to succeed, as we have many Windows TCC users in the LLM space, and they all hit this issue (it only took me 1 min to quickly google these):

• ERROR: Failed to create instance: unexpected error when creating modelInstanceState triton-inference-server/tensorrtllm_backend#328
• CUDA runtime error in cudaDeviceGetDefaultMemPool on [windows + 16GB V100] TensorRT-LLM#1821

[leofang]

#209 is related (the 3rd step we should take in the future to address this issue).

[leofang]

Another reason that TCC is important is because it's the default of GHA Windows GPU runner, e.g.:
https://github.com/cupy-ci-poc/cupy/actions/runs/12004661559/job/33459987560#step:5:15