Closes #581.

Summary

LazyGroupBy had no __iter__, so for k, g in ds.groupby([...]): fell back to __getitem__(0) and raised TypeError: Expected str or list, got int — a misleading message hiding a missing feature. ColumnExpr.__iter__ also dropped groupby context: for x in gb['col']: yielded raw values instead of pandas SeriesGroupBy (key, sub_series) pairs.

Changes

• datastore/groupby.py — LazyGroupBy adds __iter__, __len__, __contains__, get_group(name, obj=None), groups, indices. All delegate to a shared _pandas_groupby() helper that respects sort / dropna / selected_columns; single-col groupby uses scalar by to match the gb['col'] scalar-key convention. ngroups refactored onto the helper too.
• datastore/column_expr.py — __iter__ yields (key, sub_series) when _groupby_fields is set; plain ds['col'] iteration unchanged.
• __getitem__ errors — clearer TypeError for gb[<int>] (suggests iter / get_group) and gb[<other>] (suggests str / list of str).

Tests

datastore/tests/test_groupby_iteration.py — 44 mirror-pattern tests across 10 TestCases:

• single/multi-column iter with scalar / tuple keys, index preservation
• sort=True/False, dropna=True/False, as_index=False, column selection
• get_group single / multi / column-selected / missing-key / legacy obj= kwarg
• .groups / .indices for single & multi column
• len(gb) / key in gb
• datetime keys, NaN keys, empty / single-row / all-same-value inputs
• SeriesGroupBy iter + clear error for iter on computed expressions
• gb[<int|float|None>] error path coverage

All 44 pass locally. Verified no regression on adjacent column_expr / groupby test files (178 passed).

Verification

for (date, code), group in ds.groupby(['date', 'code']):
    print(date, code, len(group))   # works, mirrors pandas

for k, s in ds.groupby('cat')['v']:
    print(k, s.tolist())            # SeriesGroupBy semantics