closes #19600

There was a major performance regression in the linear models. This was due to the new use of the _incremental_mean_and_var().
In the case when normalize parameter is not set in the linear model calculations of the variance are not necessary. This PR exchanges it for np.average() in case when normalize is set to False.

Performance (current main):

zoomed into _preprocess_data at _base.py:

Performance (this PR):

zoomed into _preprocess_data at _base.py:

The performance is measured using the code of @jeremiedbb :

from sklearn.linear_model import ElasticNet 
from asv_benchmarks.benchmarks.datasets import _synth_regression_dataset 
data = _synth_regression_dataset(n_samples=5000, n_features=10000) 
X, _, y, _ = data 
estimator = ElasticNet(precompute=False, alpha=100, random_state=0) 
%load_ext snakeviz 
%snakeviz estimator.fit(X, y)

cc @ogrisel @jeremiedbb @agramfort