vLLM adds Gemma 4 support with optimized routing kernels

vLLM has implemented support for Gemma 4, including a change to skip CUDA-graph padding rows in the routing kernel. This optimization targets improved inference efficiency for the new model architecture.