feat: add LTX and MiniMax H3 Tiny VAE live previews

[GOvEy1nw]

Summary

Adds opt-in Tiny VAE live previews for LTX 2.3 / LTX 22B and MiniMax H3 FL2VA/Ref2VA, including pruned H3 variants. LTX 2.5 Dev and Distilled are also enabled provisionally through the existing taeltx2_3 decoder.

• Publish denoised/x0 previews early without changing sampler evolution; apply source conditioning only to a preview clone.
• Use the shared architecture/capability registry, bounded coordinator, scheduling, OOM/CPU recovery, and configuration/API/CLI/WebUI/MCP transport.
• Prefer H.264/NVENC multi-frame previews with animated WebP and static-image fallbacks.
• Preserve playback position across preview updates, isolate generations, and cancel/close preview sessions on abort or completion.
• Retain RGB fallback when a decoder is unavailable or preview decoding fails.

Decoders and compatibility

• LTX uses the pinned 128-channel temporal TAEHV decoder with MIT attribution.
• H3 uses Kijai's pinned 24-channel flat 2D decoder with Apache-2.0 attribution.
• LTX 2.5 reuses the shared LTX adapter and decoder; live preview quality for LTX 2.5 remains unverified. No new weights or sampler changes are introduced for this extension.

Validation

• python -m unittest tests.test_preview_subsystem — 27 tests run, 26 passed and one opt-in decoder fixture skipped. Coverage includes both LTX 2.5 defaults resolving to the intended decoder.
• git diff --check — passed.
• Earlier checks verified pinned decoder loading/CPU decoding and live LTX playback. No new live GPU generation was run during this update; full H3 and LTX 2.5 generation validation remains outstanding.