Add GPU memory requirements to model catalog

[anfredette]

The model catalog should include GPU memory requirements for each model to support capacity planning and feasibility checks.

Acceptance Criteria

• Calculate or document GPU memory requirements for each model in the catalog
• Account for tensor parallelism configurations (e.g., memory per GPU for TP=2)
• Add gpu_memory_gb field to data/model_catalog.json
• Update Recommendation Engine to filter out infeasible GPU types based on memory constraints
• Validate memory estimates against real deployments

Notes

• Memory requirements vary by model size, quantization, and tensor parallelism
• This prevents recommending invalid configurations (e.g., 70B model on single L4)