Add more optimizer dependency support

[gesen2egee]

Summary

• Add optimizer dependency updates for additional optimizer support.
• Pin pytorch-optimizer to commit 3d08fa0.
• Update adv-optm to 2.5.2.
• Preserve Prodigy / DAdapt learning-rate logging while supporting effective_lr logging from upstream.

[kohya-ss]

Thanks for the PR. Could you clarify why schedulefree and pytorch-optimizer need to be installed from Git commits instead of PyPI releases? For default requirements, we generally prefer released version pins for reproducibility and install stability.

Also, torch-optimi and adv-optm appear to be newly added default dependencies, but this PR does not seem to add direct code paths that require them. Since arbitrary optimizer classes can already be used by installing optional packages separately, could you remove these from requirements.txt unless they are required for a default/supported path?

[gesen2egee]

Thanks for the feedback. Here is a quick clarification:
schedulefree & pytorch-optimizer (Git commits): Because the latest PyPI releases lack new features—specifically, the Schedule-Free+ variants and several newly added optimizers.

torch-optimi & adv-optm: They provide Stochastic Rounding and Kahan Rounding implementations for common optimizers (which are fully supported and utilized by frameworks like onetrainer). They are essential for preventing gradient underflow in low-precision training