feat(minimax_h3): Text-to-Image mode with multi-seed stills

[Arturogpj]

Summary

• Adds Text to Image mode for all MiniMax H3 architectures (FL2VA / Ref2VA + pruned 20B).
• Image mode runs the shortest legal H3 clip (5 frames), keeps the first frame as the still, and skips audio export.
• Number of Images runs independent multi-seed generations (true different takes), not frames from one clip.
• Fixes VAE decode crash on the 5-frame minimum (decoded 0 frames, expected 5).
• Fixes v2i batch control stuck non-interactive after Video→Image form refresh.
• Records per-image seeds in metadata/filenames for multi-still batches.

Motivation

H3 is strong for multi-reference identity work. Running it as short stills (with optional refs on Ref2VA) makes that usable without a full video render. Community workflows (e.g. multi-ref still sheets) already treat H3 this way in Comfy; this brings a first-class path into WanGP.

Behavior

Mode	Behavior
Text to Video	Unchanged (default min length stays video-oriented)
Text to Image	5-frame clip → keep frame 0 → JPG gallery
Number of Images = N	N separate short gens with seeds S, S+1, …

Model coverage

• minimax_h3_fl2va / minimax_h3_fl2va_pruned — text / start-image stills
• minimax_h3_ref2va / minimax_h3_ref2va_pruned — multi-ref stills

Implementation notes

• Handler: v2i_switch_supported, frames_minimum_image: 5, image_batch_size_max: 8, image_mode_keep_all_frames: false
• Pipeline: image-mode path, multi-seed loop, returns seeds list for metadata
• VAE: minimum-length decode path for 2 latent tokens (5 pixel frames)
• wgp.py: honors frames_minimum_image; batch control always interactive=True (Radio when max≤8, Slider otherwise) so v2i refresh does not leave controls disabled
• media_recording: per-image seed when saving multi-image batches

Test plan

• H3 Ref2VA → Text to Image → single still succeeds
• H3 Ref2VA → Number of Images = 2–3 → distinct stills
• Metadata/filename seeds differ per still (S, S+1, …)
• No more VAE decoded 0 frames, expected 5
• Batch control clickable after switching Video → Image
• H3 FL2VA Text to Image smoke
• H3 Text to Video still works (length/audio unchanged)
• Non-H3 image models (e.g. Z-Image/Flux) batch control still OK

Discord-friendly blurb

See PR comment / release note style summary in the discussion if useful for #announcements.