[FEA]: Add a MemoryResource which uses CUDA VMM APIs to allocate memory #967

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct

Area

cuda.core

Is your feature request related to a problem? Please describe.

I would like to be able to use the equivalent of cuMemCreate, cuMemMap, and friends via a cuda.core MemoryResource

Describe the solution you'd like

I'd like to have a VMMAllocatedMemoryResource which I can create on a Device() for which allocate() will use the cuMem*** driver APIs to create memory.

Describe alternatives you've considered

Currently the only alternative is to use the bindings APIs directly. Since the cuMem*** functions are synchronous, there's no way to fit this with the MemPool APIs as-is (this is my current understanding, at least.)

Additional context

This is useful to support NVSHMEM/NCCL external buffer registration, or for more interesting cases like growing allocations without changing pointer addresses, or EGM on Grace-Hopper or Grace-Blackwell systems.