llama.cpp

sycl: use async memory allocation to fix crashes during graph recording

#16644

Merged