feat(cache): add resumable cache management

[Minamiyama]

Background

Model download progress, cached models, and virtual environments were previously managed through separate entry points. Interrupted downloads also lacked persistent pause, resume, and cleanup capabilities.

Changes

Cache Management Page

• Added a new Cache Management page under Model Management.
• Unified active, incomplete, and completed model downloads in the Model Cache tab.
• Displayed model format, parameter size, quantization, status, progress, cache paths, and Worker information.
• Added collapsible download-file details with animation.
• Standardized action buttons as icons with tooltips.
• Added a separate tab for virtual environment management.

Download Task Management

• Added support for downloading model artifacts without launching a model instance.
• Added pause, resume, cancel, and delete operations.
• Persisted unfinished tasks so interrupted downloads remain resumable after a Supervisor or Worker restart.
• Stored task state in SQLite using WAL mode.
• Added XINFERENCE_DOWNLOAD_TASK_DB_PATH for configuring the database path.

Safe Cache Cleanup

• Tracked primary, draft, ControlNet, GGUF, and Lightning repositories used by each task.
• Protected shared repositories used by multiple Workers on the same host.
• Preserved task state when cancellation times out instead of reporting success or deleting the record.
• Prevented failed repository resolution or database writes from leaving stale running task IDs.

API

Added or extended:

• POST /v1/cache/models
• GET /v1/cache/models/{cache_uid}/progress
• POST /v1/cache/models/{cache_uid}/cancel
• GET /v1/downloads
• POST /v1/downloads/{cache_uid}/pause
• POST /v1/downloads/{cache_uid}/resume
• DELETE /v1/downloads/{cache_uid}

Compatibility

• Existing model launch APIs remain unchanged.
• Pause and resume apply only to download-only tasks.
• The default database path is $XINFERENCE_HOME/download_tasks.db.
• Byte-range continuation depends on the selected model Hub SDK. If unsupported, the incomplete file is downloaded again.

[gemini-code-assist]

Code Review

This pull request introduces a new 'Download without launching' (cache-only) feature to Xinference, allowing users to download model artifacts without deploying them. It adds several REST endpoints for caching, pausing, resuming, and deleting downloads, along with a dedicated Cache Management page in the frontend. It also updates the permission guard, sidebar, and launch dialog to support this feature. Feedback on the frontend implementation highlights an issue with the polling mechanism in the Cache Management page, recommending a transition from window.setInterval to window.setTimeout with a local cancellation flag to prevent overlapping requests and memory leaks upon unmounting.

[qinxuye]

LGTM