Serve models with vLLM

[miguelgrc]

Description

We should serve the main model to be queried from inspire with vLLM. This has some advantages over Ollama, like higher throughput and out-of-the-box metrics

Work involved

• vLLM is already installed in the AI machine
• Serve the preferred model via vLLM
• Modify the AI backend to query this new port and model

Acceptance criteria

• The model queried from inspire search is running in vLLM