[Bugfix] Set separate CFG flag in Helios for CacheDiT

[alex-jw-brooks]

Purpose

Related: #2527

It looks like Helios should have has_separate_cfg=True for Cache-DiT, since we always handle +/- embeds separately, both for standard cfg & cfg zero* (here). You can also see this is the case in the Cache DiT adapter for Helios, here.

Currently, the output with cache DiT on using some models, e.g., helios base, will come out fuzzy:

For example, ground truth:

seed_131_gt.mp4

Example with Cache DiT on:

helios_cache_bug.mp4

Result after fix:

post_fix.mp4

Test Plan

Added Helios to explicit checks for has CFG=True (which were previously added for the same issue on Longcat/ltx2 #2860)

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[hsliuustc0106]

could we make this model real-time streaming output?

[hsliuustc0106]

I am trying to optimizing this model

[alex-jw-brooks]

@hsliuustc0106 Yes, I think Helios should be able to support real-time, at least if it's one request at a time, since it was originally supposed to get up to ~19 FPS on H100!

This PR is just a correctness fix though since the outputs get destroyed by not separating CFG correctly

[hsliuustc0106]

LGTM - well-justified fix with visual evidence and test coverage.

[hsliuustc0106]

@hsliuustc0106 Yes, I think Helios should be able to support real-time, at least if it's one request at a time, since it was originally supposed to get up to ~19 FPS on H100!

This PR is just a correctness fix though since the outputs get destroyed by not separating CFG correctly

please check #3737