feat: infrastructure for both local GPU setup and Modal

[ArnavBharti]

This has two scripts vllm_inference_modal.py and vllm_inference_local.py.

The Modal scripts sets up the vLLM server on A100 40GB (configurable). The local script is works on NVIDIA RTX 6000 Ada with CUDA version 12.4 (BITS Server).

Both scripts were tested with Qwen/Qwen3-Embedding-0.6B.

Modal deployment was tested using examples in README.

First time setup involves setting up Modal Volumes and caching. Subsequent deployments are fast. HuggingFace model is fixed to a commit rather than main.

After volumes are cached, response time on Modal:

[ArnavBharti]

This PR now contains code for the entire program flow.
It accepts JSON request -> converts to embeddings -> similarity search using FAISS -> respond with dish name in JSON format.

Tested on both Modal and GPU server.
Right now this runs Qwen 0.6B embedding model which is text only.