Fix automatic oversubscription scheduling with gradient accumulation

[hjinnkim]

Dependency

Depends on #2897 for occurrence-aware sampler and resume semantics. This branch is stacked on #2897 and will be rebased after it merges.

Additional issues found

• Automatic oversubscription increased the repeat count without constructing the repeated schedule, so gradient accumulation could still leave insufficient rank-local samples.
• repeats: 0 was treated as a manual override.
• Empty buckets could enter repeat division.
• Mutating the backend repeat value could affect later splits.
• Conditioning metadata copying could discard the current rank-local source schedule and overwrite canonical target bucket state with rank-local data.

Resolution

• Keep the automatic repeat calculation local to the split.
• Build a batch-aligned cyclic schedule directly at rank-local size.
• Treat only positive configured repeats as a manual override.
• Skip empty buckets.
• Preserve the source schedule and avoid persisting rank-local target bucket indices over canonical metadata.

Regression coverage for gradient accumulation, standard DP, Context Parallel, empty buckets, manual repeats, and conditioning metadata is included.