MiniMax H3 video VAE: decoding a single latent frame (off the 17k+5 grid) produces heavy artifacts — pad-to-2-tokens works; suggest graceful handling or validation

## Summary

The MiniMax H3 video VAE **decode** path produces visibly degraded output — 256px tile seams and banding in its default tiled mode, and strong ViT patch-grid artifacts untiled. The **encode** path is fine.

## Repro (master a464ac33, single 1024x1024 image, `minimax_h3_video_vae_fp16.safetensors`)

`LoadImage -> VAEEncode -> VAEDecode -> SaveImage`, compare output to input (mean abs pixel error, 0-255):

| path | roundtrip error |
|---|---|
| reference implementation (see below) | **4.6** |
| ComfyUI decode, tiled (default `tiling=True`, 256/64) | **31.4** — visible tile seams + banding |
| ComfyUI decode, `tiling=False` | **93.7** — 16px patch-grid artifacts |

Errors are dtype-independent (fp16 / `--bf16-vae` / `--fp32-vae` within 0.1 of each other), so it's algorithmic, not precision.

**Encode is healthy**: comfy-encoded latents match the reference encoder at cosine 0.9997, and decoding a comfy-encoded latent through the reference decoder gives the same 4.6 as a full reference roundtrip.

## Reference implementation

The ai-toolkit MiniMax-H3 extension carries an independent port of this VAE ([ostris/ai-toolkit `extensions_built_in/diffusion_models/minimax_h3/src/vae.py`](https://github.com/ostris/ai-toolkit)) that implements the same 256px/64px-overlap spatial windowing with linear blending and reproduces the released pipeline's output (4.6 roundtrip on the same weights file). Diffing the tile layout / per-tile RoPE origin / blend-weight math between the two implementations should localize the defect quickly — the untiled 93.7 result also suggests the ViT decoder is only valid in <=256px windows, so the tiled path is the one worth fixing.

Happy to provide the exact test script/images or help bisect. Found while building a custom node that does per-slot decodes of single-frame H3 latents, where the seams are conspicuous on flat backgrounds; the same degradation applies to every normal H3 video decode at >=512px.

## Environment
ComfyUI master `a464ac33`, Linux, RTX PRO 6000 (Blackwell), torch 2.14.0.dev+cu132, `--gpu-only`.

Start by locating the MiniMax H3 VAE decode path behind the VAEDecode entry point, then compare its tile layout, per-tile RoPE origin, and blend-weight math with extensions_built_in/diffusion_models/minimax_h3/src/vae.py from the referenced ai-toolkit implementation. Reproduce the single-frame 1024x1024 roundtrip and verify that tiled decoding removes seams and banding while preserving the reported reference error.

Written by the indexing model from the issue text.