Support reference-to-video generation through WanGP

[georgi]

The combined WanGP provider now accepts reference images and video clips without assigning them a start-frame role. Discovery advertises reference_to_video only when the model metadata supplies usable reference modes. The adapter preserves image and video order, validates supported kinds and counts, maps images to image_refs, and maps one or two videos to video_guide / video_guide2.

Mappings were checked against the Docker-pinned WanGP revision 362c3467a70e1136ceb52eec95907205a8f88543. Its reference-video modes are V-U and V+-U, correcting the design draft's assumed mode names. Generic control mode GV is excluded. Mixed references retain the image mode, and soundtrack conditioning uses the advertised K audio mode only when requested. Start-frame generation retains image_start and S.

The Docker image advertises the new capability and pins nodetool-core commit e730deb52b4f4bb9c90fb8db714a597d87c58ee9 from nodetool-ai/nodetool-core#1057. Merge that worker prerequisite first. WanGP itself remains unmodified and isolated in its existing interpreter.

Validation: 134 Python tests passed, Ruff and Ty passed, and the public-package import boundary check returned no forbidden imports. Tests consume a metadata fixture extracted and normalized from the pinned upstream source. A real TypeScript bridge → Python worker → combined adapter integration passed image-only, video-only, two-video, mixed, soundtrack on/off, chunked output, and temporary-file cleanup. The upstream session was simulated, so this does not claim a Docker build or live GPU generation.

Companion TypeScript bridge: nodetool-ai/nodetool#5698.