Fix RFM NaN handling, drop dead numpy path, align lapse classes_

[shivamlalakiya]

Summary

Three small correctness fixes and one docs fix in preprocessing/models, found while reading through _rfm.py, _lapse.py, and _forecast.py:

• RFMTransformer, behaviour change: a gift with a NaN gift_amount used to count toward frequency while being silently dropped from monetary, so the two columns described different sets of gifts. It now excludes that gift from both, with a UserWarning naming how many rows got dropped. This means frequency can come in lower than a donor's raw row count when some of their gifts have no recorded amount.
• RFMTransformer, dead code: the docstring described a numpy-array input path (falling back to x0..xn column names), but _validate_input always required donor_id/gift_date/gift_amount by name, so that path always raised anyway and never actually worked. Removed the dead branches and now raise a clear TypeError up front telling the caller a DataFrame with named columns is required. Also added a line to the docstring noting the outputs are raw R/F/M values, not scores or frozen bins.
• LapsePredictor: classes_ is now built with sklearn.utils.multiclass.unique_labels, matching every other classifier in the package (PropensityScorer, etc.) instead of a bare np.unique(y).
• FinancialForecastModel: the module docstring's title called it a "Hybrid LSTM-ARIMA" forecaster without the caveat the class docstring carries further down. Added the same "it's really LinearRegression + MLPRegressor on residuals + a hand-rolled AR(p), no actual LSTM or ARIMA" note to the title.

CHANGELOG.md has a new ### Fixed entry under [Unreleased] covering all four.

Test plan

• Added test_rfm_transformer_nan_amount_excluded_from_frequency_and_monetary in tests/test_rfm.py, confirmed it fails against the old code (frequency=2 instead of 1, no warning) and passes with the fix.
• Updated test_rfm_transformer_validation and renamed/updated test_rfm_transformer_ndarray_input_raises_missing_columns to match the new TypeError message for non-DataFrame input.
• Added tests/test_lapse.py (none existed before) with a basic fit/classes_ test and a lock-in test that classes_ matches unique_labels(y).
• make ci: 2198 passed, 30 skipped, coverage 98.35% (floor 92%), flake8 and mypy clean.
• make riskcov: risk-tier coverage 98% against the 93% floor.