crash-fix

Llama.cpp is expected to receive a fix for split mode tensor crashes on multi-GPU setups, which currently cause VRAM exhaustion every 90-120 minutes. The fix also reportedly brings a ~35% throughput improvement over layer mode.