vLLM 0.33.0 adds new model support and moves to CUDA 13.0 only

0 is out, a release the project says spans 927 commits from 389 contributors.

Breaking changes remove Medusa and the MLP speculator, and requests sent while generation is paused now return HTTP 503. An OpenAI-compatible /v1/decisions endpoint is on by default for supported models, and images also ship under the shorter vllm/vllm name.