AnyFlow: flow map distillation core

[bghira]

This pull request adds support for a new distillation method called "anyflow" across the codebase and documentation. The main changes include implementing the AnyFlowDistiller class, registering the new method in the factory and registry, and updating documentation in multiple languages to reflect the new option.

Support for the "anyflow" distillation method:

• Added a new AnyFlowDistiller class implementing the "anyflow" distillation strategy for flow-matching models, including batch preparation, loss computation, and model-specific requirements. (simpletuner/helpers/distillation/anyflow/distiller.py)
• Registered the "anyflow" method in the distillation registry and ensured it is imported in the factory for availability. (simpletuner/helpers/distillation/anyflow/__init__.py, simpletuner/helpers/distillation/factory.py) [1] [2] [3]

Documentation updates:

• Updated the --distillation_method option in the training documentation to include "anyflow" in all supported languages. (documentation/OPTIONS.md, documentation/OPTIONS.es.md, documentation/OPTIONS.hi.md, documentation/OPTIONS.ja.md, documentation/OPTIONS.pt-BR.md, documentation/OPTIONS.zh.md) [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12]

These changes enable users to select "anyflow" as a distillation method and provide the necessary implementation and documentation for its use.

[bghira]

target_anyflow.mp4

AnyFlow validation output

[bghira]

@copilot resolve the merge conflicts in this pull request

[Copilot]

@copilot resolve the merge conflicts in this pull request

Resolved the merge conflicts in commit bffef15. The two conflicted files were:

• simpletuner/helpers/models/ideogram/quantized_loading.py: took origin/main's improved _Fp8LinearScaledMm (added _dequantized_forward fallback, better dtype handling, layout-error recovery) and the Fp8Linear.forward dtype validation
• tests/test_ideogram_fp8_linear.py: took origin/main's updated test suite covering the new functionality