fix: align generated dataset columns to model's feature_names

[majiayu000]

Summary

• Add column name alignment in LLM generators to ensure generated datasets match the model's expected feature_names
• This fixes the KeyError that occurs when different detectors generate datasets with inconsistent column names

Root Cause

Different LLM-based detectors were generating datasets with different column names:

• sycophancy: query
• harmfulness: user_input
• stereotypes: user_input
• etc.

When a model specifies feature_names=["message"], the prepare_dataframe() method fails with KeyError because the generated dataset doesn't contain the expected column names.

Solution

Add _align_columns_to_feature_names() method that:

• Reorders columns if they match feature_names but in different order
• Renames columns if count matches (positional mapping)
• Handles cases with more generated columns than expected features
• Logs warnings for edge cases

Test plan

• Added unit tests for column alignment functionality
• Existing tests should continue to pass

Fixes #2213