Describe the issue linked to the documentation

Two examples for missing-values imputation use k-neighbors imputation without scaling data first.
As a result, the approaches under-perform.
The examples are:

1. https://scikit-learn.org/stable/auto_examples/impute/plot_missing_values.html#sphx-glr-auto-examples-impute-plot-missing-values-py
2. https://scikit-learn.org/stable/auto_examples/impute/plot_iterative_imputer_variants_comparison.html

In the first example, the effect is quite small, adding scaling before calling k-neighbours imputer changes MSE for the california dataset for k-NN from 0.2987 ± 0.1469 to 0.2912 ± 0.1410 and for the diabetes dataset from 3314 ± 114 to 3323 ± 90.

In the second example (comparing iterative imputations), the change is more significant: before the change, iterative imputation with k-neighbors performed worse than imputation with mean, after the scaling -- it performs better than mean imputation.

In both cases, it is a better practice to scale data before using a k-neighbors approach which is based on distances between points.

Suggest a potential alternative/fix

I will submit a patch to fix an issue.