Function memcpy_batch 

pub unsafe fn memcpy_batch(
    src_ptrs: *const *const c_void,
    dst_ptrs: *const *mut c_void,
    size_per_copy: usize,
    num_copies: usize,
    mode: MemcpyBatchMode,
    stream: cudaStream_t,
) -> cudaError_t

Batched memcpy using cudaMemcpyBatchAsync (CUDA 12.9+) and/or individual cudaMemcpyAsync.

Takes HOST arrays of src/dst pointers - no device allocation needed. Direction is auto-determined by CUDA from pointer types using cudaMemcpyDefault.

The mode parameter controls dispatch:

• MemcpyBatchMode::BatchedWithFallback: try batch API, fall back to individual copies on error
• MemcpyBatchMode::FallbackOnly: always use individual cudaMemcpyAsync loop
• MemcpyBatchMode::BatchWithoutFallback: try batch API, return error if unavailable

Safety

• src_ptrs must point to a valid array of num_copies source pointers
• dst_ptrs must point to a valid array of num_copies destination pointers
• Each source/destination pointer pair must have at least size_per_copy bytes accessible
• stream must be a valid CUDA stream handle