LTX-2 support in Candle

Lightricks just released their LTX-2 weights and inference code -- https://github.com/Lightricks/LTX-2

It would be great to have this model in Candle; it appears to be quite fast and can generate synchronized audio and video.

Are there any foreseeable architectural blockers to this implementation? Is this something that Candle would be interested in?

Start by reviewing the linked Lightricks LTX-2 repository, especially its inference code and model requirements, then compare those needs with Candle's existing architecture. The issue names no Candle files or tests; done would require a defined implementation scope and working LTX-2 support, including synchronized audio and video generation.

Written by the indexing model from the issue text.