llama.cpp

fix: OOB reads in UGM tokenizer (precompiled_charsmap handling)

#18750

Merged