llama.cpp

Avoid fp32->fp16->fp32 conversion on cdna in ggml_cuda_op_mul_mat_cublas

#11356

Merged