Implement count-heuristic ZB-H2 pipeline schedule

[insujang]

Summary

• split OneForwardOneBackwardSchedule into schedule_1f1b.py while preserving compatibility re-exports
• add count-heuristic ZB-H2 with the u_s = 2(p-s)-1 warmup and independent count-selected F/B/W operations
• add genuine deferred linear weight-gradient computation for tied weights, DTensor TP paths, and stacked expert weights
• add ParallelConfig.pipeline_schedule dispatch and reject inconsistent per-module selections
• preserve the multimodal encoder-to-LM seam and intentionally omit ZBV/VPP

Validation

• Ruff checks passed
• 22 focused schedule, configuration, and pipeline tests passed
• full distributed parallelism integration matrix: 16 passed
• existing 1F1B multimodal seam tests plus six-rank ZB-H2 CP seam parity: 3 passed
• strict two-rank 1F1B versus ZB-H2 loss and local-gradient parity, including consecutive-step cleanup
• PP+TP and PP+EP Gloo coverage