llama.cpp b9966 caches regex patterns to reduce CPU usage on sm-tensor

The llama.cpp project released build b9966, which addresses a performance issue in the -sm tensor feature. The update removes 29 regex recompilations per tensor per token that were previously occurring on the decode thread.

• The code previously rebuilt 29 regex patterns from scratch for every tensor of every token.
• The fix caches these patterns so they are built once and reused.
• This change reduces wasted CPU cycles on the decode thread without altering behavior.