llama.cpp

Fix data race in CUDA's "cpy" kernel (influences GGML's DUP, CONT operations).

#20507

Merged