Reference Issues/PRs

Closes #5651

What does this implement/fix? Explain your changes.

This PR:

1. More intelligently chooses between np.mean and np.nanmean. The nan mask is always computed when calling np.nanmean
2. Computes the correction first and then the new normalized variance using a temp variable.

Any other comments?

With the script in #5651:

import numpy as np
from sklearn.preprocessing import StandardScaler


shape = (2 ** 16, 2 ** 12)
big = np.random.RandomState(0).uniform(-1, 1, size=shape)
scaler = StandardScaler().fit(big)

And running on main and this PR:

mprof run standard_scaler_bench.py

I get the following plot:

Where the black line is this PR and the blue line is main. We see that this PR reduces the compute time and memory usage.

CC @jeremiedbb