chore(deps): bump vllm to 0.16.0

[mplatzer]

Summary

• bump vllm in gpu optional dependency from 0.12.0 to 0.16.0
• refresh uv.lock to resolve new vLLM transitive dependency set
• update stale compatibility comment for transformers

Validation

• uv lock
• uv lock --check

---

Note

Medium Risk
Medium risk due to a major runtime dependency bump for GPU inference (vllm) plus a large lockfile refresh that changes key ML stack versions (Torch family, protobuf) and adds new transitive networking/runtime deps.

Overview
Bumps the optional gpu extra to vllm==0.16.0 and updates the transformers compatibility comment accordingly.

Refreshes uv.lock to match the new vLLM dependency graph, including new transitive packages (notably grpcio, grpcio-reflection, mcp, ijson, httpx-sse, pyjwt, sse-starlette) and several version/marker shifts such as torch/torchaudio/torchvision platform-specific pins, xgrammar updates for macOS/Linux, and a protobuf version change.

^{Written by Cursor Bugbot for commit ce29ca6. This will update automatically on new commits. Configure here.}