[FEA] cuda.core: support stream batch memory operations (incl. CUDA 13.1 atomic reductions)

Summary

CUDA 13.1 extends the stream batch memory operations with atomic reductions: a new op type
CU_STREAM_MEM_OP_ATOMIC_REDUCTION with params struct CUstreamMemOpAtomicReductionParams
(operation, flags, reductionOp, dataType, address, value, alias), plus enums
CUstreamAtomicReductionOpType (ADD/AND/OR) and CUstreamAtomicReductionDataType
(UNSIGNED_32/UNSIGNED_64), joined into CUstreamBatchMemOpParams_union.atomicReduction.
Device support is advertised by CU_DEVICE_ATTRIBUTE_ATOMIC_REDUCTION_SUPPORTED (13.1, see
also #2361).

cuda.core has no batch mem-op surface at all today — the only mentions are a
graph debug-dot print flag
(https://github.com/NVIDIA/cuda-python/blob/c000331de6c37aa4565af74b001271ffcf6d5c99/cuda_core/cuda/core/graph/_graph_builder.pyx#L123)
and a DeviceProperties support query
(https://github.com/NVIDIA/cuda-python/blob/c000331de6c37aa4565af74b001271ffcf6d5c99/cuda_core/cuda/core/_device.pyx#L803).
So this issue covers designing the base surface (the pre-12.8 wait/write-value ops) and
including the 13.1 atomic-reduction op in it — adding the new op alone is not possible.

Underlying C APIs to cover

Design sketch (draft — needs design-meeting review)

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

stream.batch_mem_op([                          # names TBD
    WaitValue(addr, value, condition=">="),
    WriteValue(addr, value),
    AtomicReduction(address=addr, value=1, op="add", dtype="u64"),  # 13.1+
])

Open questions for the meeting:

1. Scope: stream-level only, or also the graph batch-mem-op node type?
2. Typed per-op dataclasses (above) vs. lower-level pass-through of binding structs.
3. Semantics/purpose of the alias field (clarify against driver docs during design).
4. Version/device gating (op availability differs across 12.x/13.x and devices).

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