Remove VLLM_USE_PRECOMPILED/VLLM_PRECOMPILED_WHEEL_COMMIT workaround once vllm releases with transformers v5

## Context

We currently use a VCS (git fork) install of vllm in `pyproject.toml` because there is no stable vllm release yet with transformers v5 support. Building vllm from a fork means vllm's `setup.py` cannot determine the base commit on `main` and falls back to the "nightly" precompiled wheel — a moving target that may not have wheels for all architectures at any given time.

To work around this, we explicitly set `VLLM_USE_PRECOMPILED=1` and `VLLM_PRECOMPILED_WHEEL_COMMIT` (pinned to the v0.16.0 tag commit) in two places:
- `pyproject.toml` — `[tool.uv] extra-build-variables` for build-time `uv sync`
- `docker/Dockerfile` — `ENV` lines for runtime ray worker venv rebuilds

## What to do

Once vllm releases a stable version with transformers v5 support (likely v0.17):
1. Switch from the VCS git install to a pinned release version in `pyproject.toml`
2. Remove `VLLM_USE_PRECOMPILED` and `VLLM_PRECOMPILED_WHEEL_COMMIT` from `extra-build-variables` in `pyproject.toml`
3. Remove the corresponding `ENV` lines from `docker/Dockerfile`
4. Update `tools/build-custom-vllm.sh` if the precompiled wheel URLs need refreshing

## Files to update
- `pyproject.toml` (search for `VLLM_PRECOMPILED_WHEEL_COMMIT`)
- `docker/Dockerfile` (search for `VLLM_PRECOMPILED_WHEEL_COMMIT`)