llama.cpp

[SYCL] Use native subgroup size for K-quant DMMV kernels on Intel

#21700

Merged