llama.cpp

ggml webgpu: actually add softmax, fix rms_norm offset

#16400

Merged