Add opt-in fresh DiT adaLN-Zero initialization

[JiamingPan]

Adds an opt-in model.initialization: adaln_zero policy for freshly constructed native DiTs.

• Zero block conditioning projections (norm1.linear), final conditioning (proj_out_1), and prediction head (proj_out_2).
• Leave attention, FFN, embeddings, architecture, loss, and samplers unchanged.
• Preserve native initialization by default; never reapply initialization when loading or resuming checkpoints.
• Validate the full projection API before changing weights and reject unsupported models or repeated initialization.
• Document usage and add 19 regression tests.

Validation: all 97 repository tests pass locally on CPU with diffusers 0.38.0, PyTorch 2.5.1, and Python 3.12.7, including identity blocks, trainable output heads, and checkpoint preservation.

This ports the zero-modulation/output policy tested in the L16 ablation, not the complete reference DiT recipe or a guarantee of generalization.