Use 256-pixel VideoVAE tiles by default

[ArthurOstapenko]

Summary

Use 256-pixel VideoVAE tiles by default, matching vae_tile_size in the
released MiniMax-H3 config. This removes the geometry-based 256–320 default
search. Existing valid H3_VAE_TILE_PIXELS overrides remain unchanged.

Reason

At 576x1024, the previous selector chooses 320-pixel tiles. In one Ref2VA
render decoded with that plan, the frames showed a visible 16-pixel grid.

Re-decoding the same saved post-denoising F32 latents with 256-pixel tiles
reduced the mean edge ratio at 16-pixel boundaries from 1.2385 to 1.1909
horizontally and from 1.1364 to 1.0890 vertically. The raw PPM and H.264
outputs remained close at 45.36 dB PSNR and 0.987 SSIM, so this comparison
did not indicate that H.264 introduced the grid.

Tradeoff

For this 576x1024 decode on an M5 Max, the VAE took 95.13 seconds with
256-pixel tiles and 83.75 seconds with 320-pixel tiles. Peak live storage
was 9.37 GiB and 9.55 GiB, respectively. Other canvas geometries can produce
different tile-count and performance changes.

Validation

• Built h3, h3_tests, h3_semantic_vae_test, and h3_real_video_vae_test.
• Passed 1,768 checks in the generic test suite.
• Completed a 576x1024 Ref2VA render with 124 frames and 20 denoising steps.
• Repeated the 256-versus-320 comparison from one saved latent artifact.