sd: sync to master-652-92dc726

[wbruna]

LongCat-Image support, and an LTX fix.

[wbruna]

Updated to master-652-92dc726, and added Lens support (without splitting the vocab files).

Edit: I think I've figured out how to split the vocab and merge files, but I needed a small change to otherarch/sdcpp/sdcpp_vocab_extract.py.

[LostRuins]

I was able to load and generate for both models fine, although lens does not seem to work with TAE flux2 for some reason, even after I tried hooking it up to the flux2 tae. Anyway I'll merge it first and fix up the cmake.

[wbruna]

lens does not seem to work with TAE flux2 for some reason

That's odd; it should use it automatically, through upstream's sd_version_uses_flux2_vae.

[wbruna]

It's an upstream issue:

[DEBUG] model.cpp:937 - loading tensors from ./taef2.safetensors
[ERROR] model.cpp:1191 - tensor 'tae.decoder.layers.0.weight' has wrong shape in model file: got [3, 3, 32, 64], expected [3, 3, 16, 64]
[WARN ] model.cpp:991 - process tensor failed: 'tae.decoder.layers.0.weight'