vllm

[docker] Build CUDA kernels in separate Docker stage for faster rebuilds

#29452

Merged