feat(chat-app): improve model/mmproj asset-level cache UX

[leehack]

Goal

Make the chat app UI expose model and multimodal projector (mmproj) cache state as independent assets.

Current state

The current PR makes model and mmproj sources independent internally and keeps the existing Manage Models flow functional. This issue is UX polish, not a blocker for the current implementation.

Proposed UX improvements

• Show model and mmproj availability separately.
• Allow re-downloading only the missing/corrupt asset.
• Make delete/clear behavior explicit when a profile uses a shared asset.
• Explain platform restrictions clearly, especially web rejecting local filesystem paths.
• Surface capability mismatch warnings, e.g. profile declares vision but runtime projector does not report supportsVision.

Acceptance criteria

• Manage Models distinguishes model vs projector cache state.
• Missing projector state does not look like the entire model is missing.
• Web local-path rejection is user-readable.
• Capability mismatch has a clear warning/info message.
• Existing model selection/download flows remain covered by tests.

[leehack]

Second-pass PR #129 review also noted two UX/privacy follow-ups that fit this issue:

• Web fallback mode may only HEAD-verify and record markers when bridge cache prefetch is unavailable, while actual loading still uses direct URL loading. The UI/docs should avoid overstating persistence in that fallback path.
• Custom model entries can persist raw user-entered url / mmprojUrl values in SharedPreferences. This may be acceptable for an example app, but the UI should warn users before saving signed URLs/tokens or eventually separate token handling from persisted source identity.