Remove Diffusers group offload integration

[bghira]

Summary

• remove the Diffusers group-offload backend, CLI options, trainer setup, and model-specific integration
• remove group-offload controls from the WebUI, memory presets, field registry, and configuration defaults
• simplify temporary pipeline offload restoration to the supported model and sequential CPU-offload states
• remove obsolete tests and documentation across all maintained translations
• retain model CPU offload, sequential CPU offload, RamTorch, and Musubi block swap

Rationale

SimpleTuner already has weight-streaming implementations with explicit training lifecycle, adapter-residency, quantization, and distributed handling. The Diffusers group-offload path duplicated that responsibility through pipeline hooks and required model-family exceptions throughout otherwise generic loading and validation code. Removing it eliminates the overlapping state machine and its configuration surface.

Validation

• .venv/bin/python -m unittest -v tests.test_group_offload_removed tests.test_krea2_model tests.test_ramtorch tests.test_quanto_workarounds tests.test_model_field_registry tests.test_config_field_validation
• npm test -- --runInBand tests/js/training_wizard.test.js
• 89 Python tests passed with 10 unavailable-accelerator skips; 5 JavaScript tests passed
• translated documentation fence and configuration-example checks passed