refactor: replace hand-rolled columnar helpers with mloda's public ones

[TKaltofen]

Closes #82.

mloda 0.10.0 ships columnar_to_rows and homogenize_rows in python_dict_utils, so the local rag_integration/feature_groups/columnar.py (added in #81 when 0.9.0 had no public inverse of rows_to_columnar) is retired. All 27 importing files, plus demo.ipynb, now use the mloda helpers.

Strictness change

mloda's columnar_to_rows raises ValueError on any non-columnar-dict input, where the local helper passed lists through and mapped None/str to []. Per call site:

• Feature groups, connector contract tests, CLI demos: always receive a genuine columnar dict from the PythonDict framework, so the strict helper is a drop-in. A malformed partition now raises instead of silently yielding [], which is the better failure mode.
• tests/integration/helpers.py::flatten_result: keeps its tolerant shape dispatch locally (row-wise passthrough, empty fallback) and only hands genuine columnar dicts to the strict helper. Behavior is preserved branch by branch; the new tests/integration/test_helpers.py pins that dispatch.
• tests/feature_groups/test_columnar.py: retired with the module; the pivot semantics are pinned upstream in mloda.

Test and annotation cleanup

Unit tests that handed calculate_feature row-wise lists (dedup base, both retrieval evaluators, graph source chaining) now build the columnar shape the framework actually delivers, via rows_to_columnar + homogenize_rows where rows are heterogeneous. calculate_feature overrides annotated data: List[Dict[str, Any]] are widened to data: Any, matching the runtime contract and the other overrides in the repo.

Verification

• tox green (707 tests, ruff format, ruff check, mypy --strict, bandit).
• Dual review pass (independent Claude Opus agent + codex): one finding each, both addressed (stale demo.ipynb import; dedup test name clarified). No correctness findings.