Fix predict_as_dataframe: single-pass, shuffle-safe, multi-dim

[sevmag]

Closes #880

Summary

• Refactor EasySyntax.predict_as_dataframe so additional_attributes are gathered inside predict_step during the same trainer pass as the model predictions, instead of in a second for batch in dataloader loop. This keeps predictions and attributes aligned by construction and removes the implicit assumption that the dataloader replays batches in the same order.
• Drop the SequentialSampler guard: with attributes now collected in-pass, a shuffled dataloader is safe to use with additional_attributes.
• Add additional_attributes parameter to EasySyntax.predict so callers that don't need a DataFrame can still get aligned attributes back; the return type becomes List[Union[Tensor, np.ndarray]] (task tensors first, attribute arrays after).
• Support multi-dimensional attributes (e.g. direction with x/y/z components): a 2-D attribute array is flattened to one column per component (<name>_0, <name>_1, ...) so the resulting DataFrame stays tabular.
• Preserve the pulse-level repeat behavior (np.repeat(value, batch.n_pulses)) and the warn-and-skip path for length mismatches.
• Add tests/models/test_easy_model.py covering event-level, pulse-level, multi-dim, and shuffled-dataloader cases.
• Drop trailing whitespace in .github/workflows/build.yml (no behavior change).