Fix cross-execution-space error: remove CUTLASS_HOST_DEVICE from CudaHostAdapter::memsetDevice

[alexngUNC]

Summary:

• Remove CUTLASS_HOST_DEVICE from CudaHostAdapter::memsetDevice in include/cutlass/cuda_host_adapter.hpp.
• memsetDevice calls the pure virtual host method memsetDeviceImpl, which implementations use to dispatch to cudaMemsetAsync / cuMemsetD*Async. It must remain host-only.
• Fixes compile failures under --Werror cross-execution-space-call when memsetDevice is instantiated from __host__ __device__ code paths.

Context:

• memsetDevice was host-only in 3.5.0 (#1411). In 3.5.1 (#1623), CUTLASS_HOST_DEVICE was added to copy/move/empty/size so CudaHostAdapter could be used in __host__ __device__ code. memsetDevice received the same macro by mistake even though it dispatches to virtual host APIs (memsetDeviceImpl → cudaMemsetAsync / cuMemsetD*Async).
• Callers (zero_workspace, fill_workspace in workspace.h) are host-only.

Error observed when compiling with --Werror cross-execution-space-call:

error: calling a `__host__` function("memsetDeviceImpl")
from a `__host__ __device__` function("memsetDevice") is not allowed

Fixes downstream build failures (e.g. PyTorch rowwise FP8 CUTLASS on Windows with --Werror cross-execution-space-call).

cc @nkhasbag-nv