Handle None/NaN/pd.NA in primitives (#25)

[matthewhorridge]

Summary

Fixes the long-standing crash when primitives encounter missing values. Closes #25.

Before:

```python
Scale(2.0).transform(None) # TypeError: unsupported operand type
df["weight_lbs"].apply(Scale(2.0).transform) # crashes mid-apply on any NaN
```

After: nulls pass through untouched.

```python
Scale(2.0).transform(None) # None
Scale(2.0).transform(float("nan")) # nan
Scale(2.0).transform(pd.NA) #
```

Design

• `isnull(value)` helper in `primitives/base.py` — recognises `None`, float `NaN`, and `pd.NA`.
• `@handle_null` decorator — for scalar primitives whose transform expects a non-null value. The decorated method is never called for a null input; the null is returned as-is. Pairs with `@support_iterable` (outer-to-inner): a list flows through `support_iterable` first, then each element through `handle_null` before reaching the real transform. Decorator order is documented in the docstring.
• Applied to 12 scalar primitives: `Bin`, `Cast`, `ConvertDate`, `ConvertUnits`, `FormatNumber`, `NormalizeText`, `Offset`, `Round`, `Scale`, `Substitute`, `Threshold`, `Truncate`.
• Not applied to `EnumToEnum` or `NormalizeBoolean` — they already model missing values via `strict`/`default` parameters and their semantics shouldn't change.
• `Reduce` and `MapEach` reject null elements with a positional error message. The rationale is documented in their docstrings: a null inside a multi-source list usually means a source column is missing for that row, and silently dropping it is exactly the kind of quiet data corruption Primitives crash on None/NaN values - need null handling #25 warned about. The CLI's existing `--on-missing` policy continues to handle the whole-column case.

Tests

`tests/test_null_handling.py` (46 new tests, all green):

• `isnull` semantics (recognised forms, rejected non-nulls)
• `@handle_null` only invokes the underlying transform for non-null values
• Every decorated primitive passes `None`, `float('nan')`, and `pd.NA` through unchanged (parameterised)
• Scale on a list containing nulls preserves the nulls in the output list
• `Reduce` and `MapEach` raise `ValueError` with the offending index for null elements
• End-to-end: `harmonize_dataset` on a pandas DataFrame containing `NaN` and `None` runs without crashing and yields nulls at the corresponding output positions

131 tests pass overall (was 85).

Backwards compatibility

Purely additive — non-null inputs behave exactly as before. The behaviour of `EnumToEnum` and `NormalizeBoolean` (which were already null-aware) is unchanged.

Test plan

• `pytest` — 131/131 pass
• `demo/harmonize_example/run_example.py` still runs end-to-end
• Reviewer to confirm the "fail loudly on null in Reduce/MapEach" choice; the alternative would be to silently skip nulls

🤖 Generated with Claude Code