Dev

[arcticoder]

install vLLM wheel from PyPI instead of building from Git

In Dockerfile.base, replace the Git-based install of vLLM@refact_v0.7.3 with:
  RUN pip install --no-cache-dir vllm==0.7.3

- Eliminates lengthy source builds and related errors
- Dramatically speeds up image creation
- Ensures a reproducible, prebuilt GPU extension from PyPI

Revert to building vLLM from GitHub with PyPI option commented - refact-server/Dockerfile.base:switched back to git-based vLLM install -- vLLM now builds from GitHub again; PyPI wheel install left as comment

[arcticoder]

Developed and tested on: Windows 11 Pro 24H2, WSL 2 (Ubuntu 24.04 LTS, Docker 28.1.1 build 4eba377), GeForce driver 576.02 (Apr 16 2025) AMD Ryzen 5 3600 6-core 3.6GHz, 32GB RAM, GeForce RTX 2060 SUPER 8GB

Though I'm currently troubleshooting "no models found" on refact-agent engine so maybe I caused that.