Fix Volcengine Seedance video references

[2973873423]

Summary

• preserve public and asset:// Seedance video inputs as video_url / reference_video content instead of extracting still frames
• upload local canvas video and audio files before submission so Volcengine receives usable public URLs
• keep video and audio previews connected when prompt edits refresh generator inputs
• automatically use reference_image roles whenever video/audio omni references are present, while retaining first_frame behavior for image-only requests
• count only submitted media when adding trusted-asset prompt indexes

Root cause

refreshGeneratorInputViews() rebuilt video-node previews from image-only references, so connecting or editing a prompt made an already-connected video appear removed. The backend then converted public video references into image frames and mixed first-frame roles with reference media, losing Seedance 2.0 omni-reference semantics.

Testing

• node tests/test_canvas_video_input_refresh.cjs
• python tests/test_volcengine_video_reference.py
• python -m py_compile main.py tests/test_volcengine_video_reference.py
• git diff --check