Numbers match cublas, but using this code leads to LLaVA outputting
nothing but white squares.

Uses some fairly disgusting preprocessor macros to get the job done
while preserving behavior when `-DGGML_USE_CUBLAS`. With a bit of
investigation into `ggml_cuda_mul_mat_mat_batched_cublas`, these can
probably be removed or simplified.

N.B. we include the source file rather than the header file in
`ggml-cuda.cu` because `llamafile/cuda.c` assumes that everything lives
in a single compilation unit.