GPU: Fix size_t/int mismatch in NPP stream context attribute query

cudaDeviceGetAttribute writes into an int*, but NppStreamContext::
nSharedMemPerBlock is size_t. Query into a local int and cast. CI
failure on the previous commit pinned all five GPU builds (cuda-11.5
gcc-10/clang-14, cuda-12.6 gcc-13/clang-19, msvc-14.29).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
1 file changed