Support dynamic model discovery across providers

[manojp99]

Problem

Hosted-provider model pickers are sourced from the static curated matrix. New models therefore require an OpenWorker release before they appear, even when a connected account can already access them. Ollama is the only provider that currently discovers models dynamically.

Proposal

Add a provider-level model discovery interface and use authenticated vendor catalogs where available, including Anthropic /v1/models, OpenAI-compatible /models, and equivalent Gemini/cloud-provider APIs.

The implementation should:

• Handle provider pagination and normalize model IDs.
• Cache results with bounded timeouts so Settings remains responsive.
• Fall back to the curated matrix when discovery is unsupported or unavailable.
• Keep the distinction between models an account can access and matrix models OpenWorker has labeled/capability-tested.
• Preserve custom model entry and hidden-model preferences.
• Expose a manual refresh and a clear stale/fallback state in the UI.

Acceptance criteria

• Connecting a supported provider surfaces models returned for that account without an app update.
• A catalog outage does not remove existing selections or block provider setup.
• Unsupported providers retain current matrix behavior.
• Tests cover pagination, cache invalidation, transient failures, and account-specific catalogs.