nvidia-smi clock locking fixes Windows multi-GPU lag

Windows remains a challenging environment for multi-GPU local LLM setups due to aggressive gaming-centric power management. Pipeline parallelization creates micro-gaps that trigger power-saving states on cards waiting for layer work, making clock stability a more frequent performance killer than PCIe bottlenecks. This optimization is essential for mixed rigs or non-NVLink setups where synchronization overhead is higher.