vLLM v0.25.1 Fixes FFmpeg Dependency and RMSNorm Quant Fusions

The latest patch release of vLLM (v0.25.1) addresses two critical issues impacting local AI deployments: it defers runtime errors for missing system FFmpeg when importing TorchCodec, preventing startup blocks even if TorchCodec isn't used; and it guards mixed-dtype allreduce RMSNorm quant fusions to ensure stable performance across different data types. These fixes are crucial for operators using vLLM with models like Qwen/Qwen3-VL-2B-Instruct, ensuring smoother model launches and avoiding potential runtime crashes.