[BUG] ZeRO-3 inference/eval dataloader bypasses SamplerDataset wrapper

[seth-hazelnutt]

Describe the bug

When running validation/test/prediction with DeepSpeed ZeRO-3, non-training dataset loaders can miss the expected .datasets attribute and sampler behavior used by the training dataloader path. This appears to break or destabilize ZeRO-3 inference/eval flows in the current OpenFold3-preview codebase.

To reproduce

1. Run validation or prediction with a non-training dataset configured under ZeRO-3.
2. Use existing DEEPSPEED zero-3 config and DataModule setup path for non-training modes.
3. Observe failures/instability around sampler/dataloader behavior before/within evaluator execution.

Expected behavior

Validation/test/prediction dataloaders should mirror the training dataset wrapper behavior and preserve sampler semantics so ZeRO-3 can execute without unexpected data-module issues.

Notes from TCE fork

Our fork fixed this by:

• wrapping validation/test/prediction datasets in SamplerDataset during DataModule.setup().
• preserving eval sampler length (epoch_len=len(dataset)) for consistent sampler semantics.

If useful, I can provide the exact patch/commit references from the fork.