llama.cpp

fix: Gemma4 tokenizer

#21343

Merged