fix(metrics): align packed sequence accuracy with ring padding

[Excelius-Wang]

Ring sequence parallelism gathers predictions and shifted labels in a per-sequence padded layout. _compute_acc still groups them with the original unpadded cu_seqlens, so sequence accuracy can attribute errors to the wrong sample or miss them entirely. For example, one correct and one incorrect packed sequence can be reported as 100% sequence accuracy.

Derive the metric boundaries from the already-padded position IDs using get_cu_seqlens_from_position_ids. This keeps the boundaries aligned with gathered tensors without changing communication, labels, or token accuracy.

Validation:

• CPU Gloo tests call the actual trainer metric method across Ring, Ring + Ulysses, Ulysses-only, and disabled-SP configurations. They compare both sequence and token accuracy with the unsharded reference across four layouts, including aligned, single-sequence, and fully masked one-token samples. Check all-correct predictions and errors at every supervised token (152 metric comparisons), using the trainer boundary helper and the normal token-metric call shape.
• Before: both Ring configurations fail; Ulysses and no-SP controls pass. After: all configurations pass.
• Related tests: 6 passed.
• pre-commit run --all-files: passed.

No full GPU training run was performed; this fixes metric computation, not training loss.