llama.cpp

ggml : require contiguous src for ROLL on CUDA and Metal

#25928

Merged