fix: impute NaN before variable-selection steps in stats.py

[drussellmrichie]

Problem

sklearn (ElasticNet) and statsmodels (OLS/VIF) raise errors when input features contain NaN values. This is triggered when a dataset uses LightGBM native NaN-handling -- e.g. sparse binary indicators like has_garage where NaN means "not recorded" -- and the variable-selection pre-pass runs before LightGBM training.

Errors seen:

• ValueError: Input X contains NaN (ElasticNet / sklearn)
• MissingDataError: exog contains inf or nans (statsmodels OLS)

Fix

Add median imputation of NaN to the top of each of the four variable-selection functions:

• calc_elastic_net_regularization
• calc_p_values_recursive_drop
• calc_t_values_recursive_drop
• calc_vif_recursive_drop

Imputation is scoped to these pre-passes only: LightGBM training still receives the real NaN values and handles them natively at each split. A UserWarning is emitted listing the affected columns. Median imputation is a neutral choice for a variable-selection screen and does not bias which variables survive the screen.

[github-actions]

Thank you for your contribution.
Please sign our CLA at the following link:
Click here to sign the CLA.
A maintainer will verify your signature and confirm it here by commenting with the following sentence:

---

I affirm that this contributor has signed the CLA

---

_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}