fix(mag_cache): correct per-transformer step count for dual-transformer pipelines

[lcheng321]

Fixes #14025

Root Cause

MagCacheConfig is applied to both transformers with num_inference_steps equal to the full pipeline step count. In dual-transformer pipelines (Wan 2.2), transformer runs only the high-noise stage and transformer_2 only the low-noise stage. Each executes a subset of the total steps, never the full count.

Three consequences:

1. _advance_step: state.step_index >= config.num_inference_steps never fires, so state never resets and calibration output is never printed.
2. Retention window: int(retention_ratio * num_inference_steps) is computed against the full step count, protecting more actual steps from skipping than intended.
3. mag_ratios indexing: step_index no longer maps to the correct denoising timestep.

Changes

src/diffusers/hooks/mag_cache.py

• Store _original_mag_ratios before interpolation so ratios can be re-interpolated to the corrected count.
• Pin _mag_cache_config on the module after apply_mag_cache so pipelines can reach it.
• Add update_mag_cache_num_steps helper.

src/diffusers/pipelines/wan/pipeline_wan.py, pipeline_wan_i2v.py, pipeline_wan_vace.py

After boundary_timestep is computed, count actual steps per transformer and correct each config before the denoising loop. No-op when MagCache is not applied.