Add an out-of-fold target encoder (OOFMeanEncoder) #1050

Is your feature request related to a problem? Please describe.

MeanEncoder replaces a category with the mean target observed for that category. When it is
estimated and applied on the same rows, every row contributes its own target to its own
encoding. With k rows per category each row is responsible for 1/k of its own encoded
value, and at k = 1 the encoding is the label itself:

So the model scores well in training and validation collapses. MeanEncoder's documentation
warns about this, but the library offers no encoder that avoids it, and the alternatives live
in other packages.

Describe the solution you'd like

A transformer that learns its mapping out of fold, so no row is encoded using its own target.
This is the same idea as category_encoders.TargetEncoder(cv=...) and CatBoost's ordered
target statistics, neither of which integrates with the feature-engine / scikit-learn pipeline
conventions used here.

Proposed API, following the conventions of the existing encoders:

from feature_engine.encoding import OOFMeanEncoder

enc = OOFMeanEncoder(
    variables=["cabin", "embarked"],
    cv=5,                 # number of folds
    smoothing=10,         # additive smoothing towards the prior
    unseen="encode",      # encode unseen categories with the prior
    shuffle=True,
    random_state=0,
    stratify=True,
)
X_train = enc.fit_transform(X_train, y_train)   # out of fold
X_test = enc.transform(X_test)                  # full-data mapping
X_train_orig = enc.inverse_transform(X_train)

• fit_transform splits the training data into cv folds, learns the mapping for each fold
  from the remaining folds, and encodes the held-out fold with it.
• fit learns the mapping from the full training set and stores it in encoder_dict_, with
  the prior in y_prior_, so transform behaves like MeanEncoder at inference time.
• variables, missing_values, ignore_format, unseen, return_empty and smoothing
  work as they do elsewhere in encoding.

Describe alternatives you've considered

• Tell users to wrap MeanEncoder in a KFold loop by hand. It works, but it is easy to get
  wrong and it is the reason the issue exists.
• Recommend category_encoders or CatBoost. Both are good, but it means a second encoder
  library in a project that already depends on feature-engine, and category_encoders has a
  different API for variables / unseen categories / returning DataFrames.

Additional context

I have an implementation with 46 tests ready, and tests/test_encoding passes in full (389
tests) including test_check_estimator_encoders.py:

• feature_engine/encoding/oof_mean_encoding.py
• tests/test_encoding/test_oof_mean_encoder.py
• docs/api_doc/encoding/OOFMeanEncoder.rst plus the encoder table and toctree entry

Happy to open the PR, or to wait if you would rather scope it differently — for example a
cv argument on MeanEncoder instead of a separate class. I am also glad to add a
docs/whats_new entry if you tell me the format you prefer for unreleased versions.