Add vLLM as a model provider

[mmargenot]

While I like Ollama a lot, I have found vLLM to be more effective as a tool when working with local LLMs. Adding vLLM would expand functionality and potentially make the embedding generation (especially in batch) more efficient.

[mmargenot]

Ollama seems less robust in recent testing, so this is a particularly helpful one.