fix: convert boundary_ratio from timestep space to index space via inverse sigma_shift

[Auraithm]

Summary

boundary_ratio=0.9 represents the DiT1/DiT2 switch point at timestep 900. In inference, this value is correctly compared against actual timestep values. However, in training_step, it was incorrectly used as an index fraction for max/min_timestep_boundary.

Due to sigma_shift (flow_match.py L57), the scheduler's timestep array is non-linearly spaced — timestep 900 sits at index ~250 (shift=3), not index 900. Using 0.9 as the index boundary caused DiT1 to be trained on 90% of indices and DiT2 on only 10%, while in inference DiT2 handles the majority of denoising (timestep 900→0).

Fix

Apply inverse sigma_shift to convert boundary_ratio from timestep space to index space before passing it to max/min_timestep_boundary. This aligns training with inference behavior.

• shift=3 → index boundary = 0.25
• shift=5 → index boundary = 0.358 (matches DiffSynth-Studio PR #806)

Reference

• fix training boundary for wan2.2 A14B modelscope/DiffSynth-Studio#806