Ulysses-SP WO A2A SAC fix

[AlbedoWang]

Summary

• Save the AutoParallel Ulysses-SP attention output A2A and post-WO boundary under GraphTrainer eager SAC.
• Prevent backward rematerialization of the A2A and WO while preserving the existing eager SAC linear counter.
• Scope the rule to the realized A2A-to-linear-to-residual graph using module FQN metadata.

Why

The row-sharded WO-to-reduce-scatter path already saves its logical output through the collective policy. The Ulysses-SP A2A-to-WO path did not have an equivalent save boundary, so backward recomputed one A2A and one WO per transformer layer.

Validation

• Focused GraphTrainer AutoParallel tests: 9 passed on the PR branch.
• Full validation on the original experiment branch: 14 AutoParallel tests passed; 202 graph-pass tests passed and 1 skipped.
• A 32-H100 paired run reduced target WO A2A calls from 64 to 32 and eliminated all 32 recomputed calls.
• Patched 4K fresh measured 287.678 ± 0.334 ms/step versus 295.419 ± 0.863 ms/step for 2K replay on the same allocation.
• Patched 4K fresh used 6.724 GiB active memory versus 6.667 GiB for 2K replay.

Known issue

Strict bitwise parity between patched and unpatched fresh runs did not pass. The step-2 relative differences were 2.94e-5 for loss and 1.29e-4 for grad norm, so the change remains draft while this numerical drift is investigated.