Startup preloading and hooks

Use startup hooks when an external action must run as llama-swap starts, and preload only models your machine can keep resident. A model with ttl: 0 stays loaded after it has been requested; preloading does not make an incompatible model fit in memory.

Keep hook commands fast and observable. A failing hook or a model that cannot become healthy should be diagnosed in logs before adding retries.