Enable dynamic timestep shifting for Flux/Flux2 presets

[dxqb]

Summary

BFL confirmed both Flux 1 and Flux 2 were trained with dynamic timestep shifting with the scheduler-derived parameters. Change presetes, and remove the "likely wrong and unknown" caveat for Flux2 from the tooltips, keeping it for Z-Image only.
Flux2 uses a higher, "empirical" shift for inference. During training, the same shifting parameters as for Flux1 were used.

It's likely that other similar models (Ernie, Z-Image, ...) were trained the same way. But we don't know, so keeping it off for other models for now.

Test plan

• pre-commit run --all-files passes
• Launched the affected UI or script and exercised the change
• Tested with at least one real preset / config when relevant (note which: ____)

AI assistance

• AI-assisted — I have read every line in this diff and can defend each change

[bghira]

they're all trained with noise levels that scale with resolution, otherwise it is not completely destroying the signal. even SDXL benefits from the change especially over 512px area

[dxqb]

they're all trained with noise levels that scale with resolution, otherwise it is not completely destroying the signal. even SDXL benefits from the change especially over 512px area

They all do, but at what parameters is the question.
Flux2 uses a much higher empirical shift during inference. Do they do that also for training? No they don't, they use the Flux1 shifting parameters. This is what this PR adds.

Ideogram4 uses a lower, resolution dependant shift.
Microsoft Lens: there was something in their techreport
Z-Image: they mention in their paper that they do, but not at what parameters. Maybe at Flux1 parameters
Others: unknown (to me)

[bghira]

do we know of a way to search for it?

[dxqb]

do we know of a way to search for it?

wdym?

[bghira]

a measurement we can make to identify ideal shift