llama.cpp merges Gemma 4 tokenizer fix

Tokenizer bugs look small, but they can quietly wreck long-session behavior and tool calling, so this is the kind of fix that materially improves real-world local inference.

• –The change is low-friction for users: the PR explicitly says GGUF files do not need to be regenerated.
• –The bug was subtle but important, involving SPE tokenization behavior and newline grouping that caused mismatches with the reference tokenizer.
• –The maintainer comments show it was validated against multiple test cases and compared with Transformers AutoTokenizer, which is the right bar for correctness.
• –For Gemma 4 users on llama.cpp, this is a reminder that pulling latest main can matter just as much as chasing new features.