[FEA] cuda.core: memory-pool location routing and new pool attributes (CUDA 13.0/13.2)

Summary

CUDA 13.0 makes memory-pool routing location-keyed: pools are addressed by
(CUmemLocation, CUmemAllocationType) — covering device pools, host / host-NUMA pools, and
managed pools (CU_MEM_ALLOCATION_TYPE_MANAGED, 13.0) under one surface:

• cuMemGetDefaultMemPool(CUmemoryPool* pool_out, CUmemLocation* location, CUmemAllocationType type)
• cuMemGetMemPool(CUmemoryPool* pool, CUmemLocation* location, CUmemAllocationType type) (current pool)
• cuMemSetMemPool(CUmemLocation* location, CUmemAllocationType type, CUmemoryPool pool)

CUDA 13.2 adds six CUmemPool_attribute members: CU_MEMPOOL_ATTR_ALLOCATION_TYPE,
EXPORT_HANDLE_TYPES, HW_DECOMPRESS_ENABLED, LOCATION_ID, LOCATION_TYPE, MAX_POOL_SIZE.

Current cuda.core state:

• cuMemGetMemPool is already used internally to obtain non-owning pool handles
  (https://github.com/NVIDIA/cuda-python/blob/c000331de6c37aa4565af74b001271ffcf6d5c99/cuda_core/cuda/core/_memory/_memory_pool.pyx#L283);
  cuMemGetDefaultMemPool / cuMemSetMemPool are unused.
• None of the 13.2 pool attributes are exposed on the memory-resource classes.
• CU_MEM_LOCATION_TYPE_NONE (13.0) is used internally and, together with
  CU_MEM_LOCATION_TYPE_INVISIBLE (13.2), is deliberately unmapped in the public StrEnum
  (https://github.com/NVIDIA/cuda-python/blob/c000331de6c37aa4565af74b001271ffcf6d5c99/cuda_core/tests/test_enum_coverage.py#L59-L61)
  — this issue should decide their public fate.

Relates to the memory-resource architecture work (#209, #528, #726) — none of which tracks
these APIs.

Underlying C APIs to cover

cuMemGetDefaultMemPool, cuMemGetMemPool, cuMemSetMemPool; CUmemPool_attribute members
ALLOCATION_TYPE, EXPORT_HANDLE_TYPES, HW_DECOMPRESS_ENABLED, LOCATION_ID,
LOCATION_TYPE, MAX_POOL_SIZE; CU_MEM_LOCATION_TYPE_{NONE,INVISIBLE} mapping decision.

Design sketch (draft — needs design-meeting review)

[!IMPORTANT]
Starting point only, not a settled design — review in the cuda.core design meeting.

• Read-only properties on DeviceMemoryResource (and the host/managed resources) for the new
  pool attributes; max_pool_size may warrant a setter — TBD.
• Device.default_memory_resource (and a host-side equivalent) backed by
  cuMemGetDefaultMemPool, returning the corresponding memory-resource wrapper.
• cuMemSetMemPool is a process-global state mutation — either expose it as an explicit,
  loudly-documented function, or deliberately keep it unexposed initially.

Open questions for the meeting:

1. Expose the global setter at all (footgun) vs. get-only in the first pass?
2. How (location, allocation type) keying maps onto cuda.core's Device/Host location
   objects (incl. host-NUMA IDs).
3. Does HW_DECOMPRESS_ENABLED belong here or with a future decompress feature?
4. Public StrEnum mapping for NONE/INVISIBLE location types.

References

• Driver docs: https://docs.nvidia.com/cuda/cuda-driver-api/
• Found during the CUDA 12.8 → 13.3 bindings vs. cuda.core gap sweep (2026-07-14)

-- Leo's bot