llama.cpp

sycl: split long rows in TOP_K instead of one workgroup per row

#27847

Merged