bugfix: checkpoint preview page missing validation samples

[bghira]

This pull request adds support for video assets in checkpoint previews, enhances the checkpoint asset management during training, and improves the dataloader UI with new features and minor bug fixes. The main themes are expanded checkpoint asset handling (including video), improved UI for dataset management, and better device handling for ramtorch and text encoders.

Checkpoint Asset Management and Video Support

• Added support for copying both image and video validation assets (e.g., .mp4, .avi, .mov, .webm) to checkpoint asset folders during checkpoint saves, and updated the logic to ensure the checkpoint preview cache is invalidated when assets change. [1] [2] [3] [4] [5] [6]
• Extended backend and frontend logic to handle video files in checkpoint previews, including correct MIME type detection, encoding, and UI display (with video controls and autoplay in previews/lightboxes). [1] [2] [3] [4] [5] [6] [7]

Dataloader and Dataset UI Improvements

• Added a "Disable dataset" checkbox footer to expanded dataset list items for easier dataset management, and included new styles for the expanded footer. [1] [2]
• Made the "Basic" tab and related controls unavailable for audio datasets, matching the logic already applied to text_embeds and image_embeds. [1] [2] [3]
• Added a "Metadata Refresh" section to the advanced dataset modal, allowing users to specify a metadata update interval for datasets (except for text_embeds and image_embeds).

Ramtorch and Device Handling

• Improved ramtorch integration by ensuring that when only a percentage of linear layers are replaced, the remaining eligible layers are moved to the correct device, and updated comments for clarity in the text encoder move logic. [1] [2] [3]

[bghira]