Fix: rebuild llama-server for CUDA 12.6 L4

Fix: use prebuilt llama-server binary, update to ClarityGuard-v2