Inconsistent behavior with bagging & base estimator class_weight

Describe the bug

Bagging Classifier transforms y to indices between 0..N but does not impact the possible base_estimator.class_weight attribute making the fit method fails when using class_weight: the class could not be found.

Steps/Code to Reproduce

This is working fine:

from sklearn.datasets import make_classification
from sklearn.linear_model import PassiveAggressiveClassifier

X, y = make_classification(n_samples=10000, n_features=5, n_redundant=0, n_clusters_per_class=1, weights=[0.5])
# let's say y is not 0 or 1 but something else like 1 and 2 for instance:
y += 1
# now i fit a classifier with class_weight, works fine
PassiveAggressiveClassifier(class_weight={1: 1, 2: 1}).fit(X, y)

Now i'm just wrapping it around BaggingClassifier:

from sklearn.ensemble import BaggingClassifier

BaggingClassifier(PassiveAggressiveClassifier(class_weight={1: 1, 2: 1})).fit(X, y)

I'm getting:

ValueError: Class label 2 not present.

It seems it comes from here : https://github.com/scikit-learn/scikit-learn/blob/95119c13af77c76e150b753485c662b7c52a41a2/sklearn/ensemble/_bagging.py#L654

Because bagging transform y but then class_weight attribute of the base estimator is not aligned with the new labels.

Is it expected or this is a bug ?

Versions

 System:
    python: 3.7.9 (default, Feb 13 2021, 00:48:23)  [GCC 10.2.1 20210110]
executable: /home/.../.pyenv/versions/3.7.9/bin/python3.7
   machine: Linux-5.10.0-3-amd64-x86_64-with-debian-bullseye-sid

Python dependencies:
          pip: 21.0.1
   setuptools: 53.0.0
      sklearn: 0.24.1
        numpy: 1.20.1
        scipy: 1.6.1
       Cython: None
       pandas: None
   matplotlib: None
       joblib: 1.0.1
threadpoolctl: 2.1.0

Built with OpenMP: True

Start in sklearn/ensemble/_bagging.py around the referenced line and reproduce the failure with BaggingClassifier wrapping PassiveAggressiveClassifier and nonzero class labels. Trace how y is transformed and how the base estimator's class_weight is used; done means the reported example fits correctly without the class-label error, with regression coverage for the behavior.