Indexed CUDA Permutation Dtype

Problem

The non-transposed raw CUDA indexed CSR kernels accepted an int64 indptr and Python aligned perm to that dtype, but the CUDA source read perm as an int32_t pointer. This interpreted 64-bit permutation entries with the wrong width and selected incorrect weight positions.

Contract

For indexed heterogeneous CSR CUDA kernels, perm.dtype must equal indptr.dtype. Both are signed int32 or int64 offset/index arrays. The CUDA kernel must read perm through the dispatched IndptrT type.

Scope

Apply the contract to the non-transposed raw CUDA MV and MM indexed kernels. The existing Python normalization remains responsible for aligning perm to indptr; the CUDA FFI entry points reject a mismatched ABI defensively.