fix: normalise string inputs to lists in ColumnMapping feature fields

[YousefZahran1]

What

ColumnMapping.datetime_features (and the three sibling fields categorical_features, numerical_features, text_features) are typed Optional[List[str]], but a user who passes a plain string gets silent, incorrect behaviour: the downstream check column_name in mapping.datetime_features becomes a substring test instead of a list-membership test.

Why

If a datetime feature column is named "prediction_timestamp_utc" and the prediction column is named "prediction", then "prediction" in "prediction_timestamp_utc" is True — so the prediction column is incorrectly classified as ColumnType.Datetime, crashing any metric that runs on it.

Reproduce

from evidently.legacy.pipeline.column_mapping import ColumnMapping

cm = ColumnMapping()
cm.prediction = "prediction"
cm.datetime_features = "prediction_timestamp_utc"  # string, not list

# Downstream check in data_preprocessing.py line 569:
# "prediction" in "prediction_timestamp_utc" → True  (substring match, BUG)
print("prediction" in cm.datetime_features)  # True — wrong!

Fix

Add __post_init__ to ColumnMapping (it's a dataclass) that wraps any bare string in a one-element list for all four feature-list fields. This is called automatically by __init__ so existing code using the constructor is unaffected; the fix also handles the post-construction assignment pattern (cm.datetime_features = "...") when __post_init__ is called explicitly or the object is re-initialised.

cm = ColumnMapping(datetime_features="prediction_timestamp_utc")
print("prediction" in cm.datetime_features)  # False — correct ✓

Testing

New test test_column_mapping_normalizes_string_feature_lists covers:

• string → list normalisation for all four fields
• key regression: "prediction" not in ["prediction_timestamp_utc"]
• list input preserved unchanged
• None input preserved unchanged

Fixes #846