Add Wan2.2 T2V-A14B diffusion training presets

[zirui]

Summary

• Adds native diffusion presets for Wan2.2-T2V-A14B high-noise and low-noise SFT stages, aligned with NeMo AutoModel's two-transformer recipe (transformer / transformer_2, boundary ratio 0.875, flow shift 3.0).
• Teaches the Wan model loader to select Diffusers transformer subfolders and remap WanTransformer3DModel state dict keys into Primus WanDiT keys.
• Extends scheduler overrides and adds focused tests for scheduler mapping, Diffusers subfolder selection, and full A14B key remapping.

Base / merge plan

This draft PR intentionally targets dev/zirui/wan (PR #779) to keep the review diff small. After #779 lands in main, this branch should be rebased/retargeted to main before final review.

Test plan

• python -m pytest tests/unit_tests/backends/diffusion
• pre-commit run --all-files
• Prepare-hook validation for both A14B examples with local /mnt/shared/zirui/models/Wan2.2-T2V-A14B-Diffusers and Wan2.1-14B encoder assets
• Full key-level smoke check: Diffusers A14B transformer keys remap exactly to a meta-initialized Primus A14B WanDiT state dict
• Add benchmark tests and align loss and performance with AutoModel.

Notes

• The A14B DiT loads from the Diffusers pipeline directory, but the native Primus Wan path still expects frozen UMT5/VAE encoder assets in original Wan .pth format. The presets default those to compatible Wan2.1-14B assets while using the A14B Diffusers tokenizer.