fix(preprocessor): make handle_missing="error" actually reject NaN

[ChrisW09]

Fixes #16.

Problem

The Preprocessor docstring promises that "error" lets missing values "reach the
transformers, which then raise on NaN". Only PLE did.

handle_missing did two things: dropped the SimpleImputer, and got forwarded as a
constructor argument to exactly one transformer (ple, via NUMERICAL_METHODS). Everything
else was left to notice NaN on its own — and doesn't. The scikit-learn scalers ignore missing
values by design, and the PreTab expansion families declare allow_nan:

hm=error minmax : no error; output all finite = False
hm=error rbf    : no error; output all finite = False
hm=error ple    : raised ValueError - Input contains NaN.

The feature-map case is worse than plain pass-through: unsupervised center placement runs
np.percentile over the column, so a single NaN makes every center NaN and the entire
feature block NaN — not just the ten affected rows.

Fix

Enforce the policy in one place rather than per transformer. get_numerical_transformer_steps
now puts a RaiseOnNaNTransformer in the slot the imputer would have occupied when
add_imputer is False:

if add_imputer:
    steps.append(("imputer", SimpleImputer(...)))
else:
    steps.append(("nan_check", RaiseOnNaNTransformer()))

This is the option the issue recommended, and it makes the documented contract literally
true. The error names the setting and its remedy:

PretabDataError: Input contains NaN, but handle_missing='error' was requested.
Fix: pass handle_missing='median' to impute missing values before the numerical
method, or remove them from the input.

Result

All 22 registered numerical methods now reject NaN under handle_missing="error" — verified
by iterating NUMERICAL_METHODS, with zero methods failing to raise. The "median" path is
untouched and still imputes.

Tests

Ten added:

• tests/test_reproducibility.py (next to the existing handle_missing tests): a
  12-method parametrised rejection test, a 3-method check that "median" still imputes, a
  check that the error is a PretabDataError naming the option, a pipeline-shape assertion
  that nan_check and imputer are mutually exclusive, and a check that clean data still
  transforms normally under "error".
• tests/test_encoder_feature_counts.py: four unit tests for RaiseOnNaNTransformer itself.

Full suite: 504 passed, 9 xfailed. pyright unchanged at 71. Ruff on pretab/ clean apart
from the pre-existing RUF100 in __init__.py — my first pass added a RUF022 by inserting
the new export out of alphabetical order in __all__, now fixed.

🤖 Generated with Claude Code