Ollama model metadata now cached by digest

When users browse available Ollama models in the picker, each refresh previously triggered fresh API calls to fetch context-window and capability metadata for every model. On large model lists, this meant waiting on dozens of redundant requests.

A new caching layer now stores metadata from Ollama's /api/show endpoint, keyed by model name and digest. Repeated picker refreshes skip straight to the cache for unchanged models. The cache holds up to 256 entries and automatically invalidates when a model's digest changes — the model was updated, so its metadata might be different too. Empty responses trigger retries before caching, ensuring partial or transient failures don't poison the cache.

This change lives in the Ollama extension's model enrichment pipeline. It does not alter model semantics — the same metadata is returned, just fetched fewer times.