Fix Arrow dataset loading with Windows workers

[kiwoongyoon]

Title: Fix Arrow dataset loading with Windows workers

Summary

• Defer GluonTS FileDataset construction until a data-loader worker starts iterating.
• Keep the existing filtering, sampling probabilities, and worker-sharding behavior unchanged while avoiding pickling PyArrow's open reader on Windows.

The supplied issue snapshot is unassigned and contains no comments or related-PR indication. Please recheck open and recently merged PRs before opening this draft.

Test evidence

• python3 -c 'import ast,pickle,sys,types; from pathlib import Path; tree=ast.parse(Path("scripts/training/train.py").read_text()); node=next(n for n in tree.body if isinstance(n,ast.ClassDef) and n.name=="LazyFileDataset"); mod=types.ModuleType("lazy_dataset_test"); mod.Path=Path; opened=[]; mod.FileDataset=lambda **kwargs: opened.append(kwargs) or ["entry"]; sys.modules[mod.__name__]=mod; exec(compile(ast.Module(body=[node],type_ignores=[]),"scripts/training/train.py","exec"),mod.__dict__); dataset=mod.LazyFileDataset(Path("sample.arrow")); assert opened==[]; restored=pickle.loads(pickle.dumps(dataset)); assert opened==[]; assert list(restored)==["entry"]; assert opened==[{"path":Path("sample.arrow"),"freq":"h"}]'
• python3 -c 'from pathlib import Path; compile(Path("scripts/training/train.py").read_text(), "scripts/training/train.py", "exec")'
• git diff --check

All commands exited 0.

Risks or notes for maintainers

• The focused assertion verifies that the production wrapper stays unopened through pickling and opens the expected dataset only on iteration.
• A native Windows training run and the full pytest suite were not available in the Linux Builder environment, which lacks pytest, torch, and gluonts.
• The change does not disable workers or reject the reported configuration.

Fixes #285