ComfyUI backend: add MiniMax H3 3D latent upscaler support to Refiner

[tonetxo]

Summary

This PR adds native support for the MiniMax H3 3D Conv Latent Upscaler (MinimaxH3LatentUpscaler3D) in SwarmUI's Refiner stage.

Changes Included

1. Installable Features (InstallableFeatures.cs):
   • Added 1-click install entry for Comfyui_Minimax_h3_latent_Upscaler (LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler).
2. Backend Feature Mapping (ComfyUIBackendExtension.cs):
   • Mapped MinimaxH3LatentUpscaler3D to the minimax_h3_latent_upscaler feature flag.
3. Workflow Generation (WorkflowGeneratorSteps.cs):
   • Joint AV Latent Handling: Since MiniMax H3 operates with joint Audio-Video flow (FLOW_AV), passing raw video latents to the second-stage sampler causes IndexError: audio_src = x[1]. We separate the joint AV latent using LTXVSeparateAVLatent, upscale the video latent with MinimaxH3LatentUpscaler3D, and recombine with LTXVConcatAVLatent.
   • Patch Alignment: Enforced 32-pixel spatial alignment ((width / 32) * 32) for MiniMax H3 to ensure patchify_video (which uses 2x2 latent patches) never fails with shape mismatch errors on arbitrary upscale ratios (e.g. 1.33x).
   • Enabled temporal chunking and force unloading on the upscaler to keep VRAM overhead minimal.

Usage Note (T2V / R2V vs Init Image I2V)

• This is intended for Text-to-Video (T2V) and Reference-to-Video (R2V / Image Prompt).
• Due to SwarmUI's general step ordering (the Refiner stage runs at step -4 prior to ImageToVideo at step 11), enabling the Refiner alongside Init Image applies the refiner to the initial 2D frame rather than post-processing the generated video. For image-guided generation with latent upscaling, users should use MiniMax H3's native R2V mode (supplying reference images in Prompt Images / ref_images) rather than Init Image.

Tested and verified on Ubuntu with RTX 50-series hardware.