twinflow: adversarial loss, doc updates

[bghira]

Closes #2452

This pull request adds support for the full self-adversarial training branch of TwinFlow, as described in the original paper, across all documentation and the core model helper code. It introduces new configuration options to enable and control adversarial and rectification losses (L_adv and L_rectify), and implements the corresponding loss computations and fake sample generation in the training pipeline. Documentation in English, Spanish, Portuguese, Hindi, Japanese, and Chinese is updated to explain these new features and settings.

TwinFlow adversarial branch support

• Added new configuration options to enable adversarial training: twinflow_adversarial_enabled, twinflow_adversarial_weight, and twinflow_rectify_weight in _twinflow_settings within simpletuner/helpers/models/common.py.
• Implemented methods in simpletuner/helpers/models/common.py for generating fake samples (_twinflow_generate_fake_samples) and computing the adversarial (_twinflow_compute_adversarial_loss) and rectification (_twinflow_compute_rectify_loss) losses, following the TwinFlow paper.

Documentation updates (all languages)

• Updated English, Spanish, Portuguese, Hindi, Japanese, and Chinese documentation (documentation/distillation/TWINFLOW.md, .es.md, .pt-BR.md, .hi.md, .ja.md, .zh.md) to describe the new adversarial branch, configuration flags, and the behavior and purpose of L_adv and L_rectify losses. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18]

These changes make it possible to optionally enable full self-adversarial training in TwinFlow, allowing for improved distribution matching and trajectory alignment without the need for an external discriminator, and ensure that documentation is clear and up-to-date in all supported languages.