GPU-shared model cold start

Design the GPU-sharing and cold-start layer for a platform that hosts thousands of customers' fine-tuned models, where most models are idle most of the time but must respond within ~2s when called, and you can't afford a dedicated GPU per model. Walk through how you pack many models onto shared GPUs, how you handle a request for a model that's currently unloaded (cold start), and how you isolate tenants so one customer's traffic or a bad model can't starve or crash others.