[Enhance] Show dataset build progress with cache status

[HAOCHENYE]

Add a tqdm progress bar to build_datasets and expose JsonlDataset.cached
so the bar can report whether each dataset hit the tokenize cache. Keep
each jsonl path bound to its own dataset/tokenize_fn config while
expanding paths, instead of reusing the last loop config across all
paths. Also fix a stale get_rank comparison (get_rank == 0) that left the
token-count debug log permanently disabled.

[HAOCHENYE]

@claude review

[claude]

Claude: The tqdm progress bar is created but never closed. When using manual bar.update(1) instead of iterating over the wrapper, you need an explicit bar.close() after the loop — otherwise the final state may not render correctly (e.g., the cursor stays on the bar line).

Suggested change

	return datasets
	if bar is not None:
	bar.close()
	return datasets

[claude]

Claude: Good catch fixing get_rank == 0 → get_rank() == 0 here. However, the same bug exists at line 402 in DataloaderConfig.build:

if self.pack_level != "none" and get_rank == 0:

Since this PR already fixes the deprecated copy, it would be good to fix the active one in DataloaderConfig.build as well — that code path is the one users actually hit.

[claude]

Claude: Review Summary

Good set of changes — the loop restructuring correctly fixes the multi-dataset config-sharing bug, the progress bar is a helpful UX improvement, and the get_rank == 0 → get_rank() == 0 fix is a real bug.

Issues found:

1. Missing bar.close() — the tqdm bar is manually updated but never closed, which can leave the terminal in an odd state.
2. Incomplete get_rank fix — the same get_rank == 0 bug (comparing the function object instead of calling it) exists at line 402 in DataloaderConfig.build. Since this PR already addresses the deprecated copy, fixing the active code path too would be worthwhile.

Both are minor fixes. LGTM otherwise.