Fixes #15845 revert #14545
Some users complain about the new check_is_fitted and I prefer to revert the change.
The core issue is that some estimators (e.g., TfidfVectorizer) set private attributes in __init__ and this seems reasonable:

from sklearn.utils.validation import check_is_fitted
from sklearn.feature_extraction.text import TfidfVectorizer
vectorizer = TfidfVectorizer()
check_is_fitted(vectorizer)
# pass

We can't simply remove v.startswith("_") because some estimators do not create any public attributes in fit (e.g., TfidfTransformer).