feat(vllm serving): Define requirements for models to be served via vLLM

[christian-pinto]

Objective

The objective of this issue is that of defining a model's requirements for serving via vLLM, in addition to the dependencies resolution.
Some of the models will support serving via vLLM. If this is the case we first need to validate this integration before we can add the model to Algorithm Nexus.

Expected outcome

Upon completion of this issue the model requirements document defined in #2 is updated with the vLLM serving requirements, and the CONTRIBUTING_NEW_MODEL.md is updated accordingly.

[christian-pinto]

For a model to work with vLLM there are the minimal requirements:

• The model architecture is present in vLLM and if not it needs to be dynamically added by means of a vLLM General Plugin
• If the model requires pre/post processing of the data before the actual inference is done, this requires for a vLLM IO Processor plugin to be provided.

Things to consider:

• Plugins need to be installed as python entrypoints
• Different distribution packages might try to register a plugin with the same name, and we need to prevent this.
• Models need to provide a test that explicitly verifies the integration is working