Restore legacy action training alongside TI2V inference

[zjn20030811]

Summary

• restore the published Wan2.1 noise, global adaln, and action-free training paths while retaining their text/image cross-attention context
• preserve the Wan2.2 TI2V action-context and temporal-modulation contract as adaln-only
• register only the active action-encoder branch, validate non-empty action checkpoints against that branch, and accept both training and inference checkpoint prefixes
• pass the explicit action configuration into the training builder and skip action statistics, paths, and loaders in action-free runs

Motivation

The current training entry point still passes an args keyword that the pipeline builder no longer accepts. The builder also configures every DiT as text-free Wan2.2 TI2V, even though the published Wan2.1 training configs use the legacy noise or global AdaLN contracts. This prevents those configurations from starting and can leave unused action parameters under DDP with find_unused_parameters: false.

This change selects the conditioning architecture from the loaded DiT, keeps the existing Wan2.2 inference behavior, and restores the legacy paths without inventing unsupported Wan2.2 modes.

Validation

• python -m pytest -q — 21 passed
• Ruff on the regression suite and fatal-rule checks on all changed Python files
• py_compile on all changed Python files
• mixed legacy checkpoints load into either selected branch, while wrong-architecture action checkpoints are rejected