Feat: Temporal tile custom size with overlap

[stduhpf]

Summary

This PR improves the temporal tiled decoding for the LTX2.3 Video VAE by adding support for custom tile sizes (processing multiple latent frames per batch) and temporal overlap/padding to ensure smoother transitions between "tiles" as the VAE decoder has access to the near future.

For now it ses env variables VAE_TILE_FRAMES and VAE_TILE_PAD to control the effect. In my experience, setting VAE_TILE_FRAMES=4 and VAE_TILE_PAD=1 seems to give very decent results (see comment).

Related Issue / Discussion

#1490

Additional Information

How it works:

Currently, when temporal tiling is enabled, it switches from decoding all latent frames at once to decoding them sequentially one-by-one. However, this causes choppy transitions every 8 frames (corresponding to 1 latent frame).

To fix this, instead of processing one frame at a time, latent frames are now decoded in small, overlapping batches (tiles). Each batch has access to the cached context of all the previous ones (like it is already the case when decoding frames sequentially). To ensure seamless transitions, the final frame(s) of a batch act as an overlap; they are discarded from the current output and re-processed at the start of the next batch.

Quick comparison

	No tiling	frame-by frame	tile = 4 , pad = 1
result	output.-.full.webm	output.-.tiling.webm	output.-.t4p1.webm
compute buffer size	14381.63 MB	2933.75 MB	10143.13 MB
decode_first_stage time	28.77s	13.64s	18.09s
tiles processed	1 (or 0?)	13	5

Checklist

• I have read and confirmed this PR follows the contribution guidelines.
• Use args instead of env variables