EHN accept l2 metric with ward linkage in agglomeration clustering

[glemaitre]

closes #18531

We should be able to use the combination metric="l2" and linkage="ward".
It is equivalent to metric="euclidean".

[ogrisel]

Are we sure we want to introduce aliases like this. We did recently the opposite for loss/objective functions were we deprecated synonyms.

[glemaitre]

Deprecating "l2" or making it an alias for "euclidean" would be fine with me.

The deprecation of "l2" would require more work to be consistent across scikit-learn.
I would expect to start deprecating wherever we have both "euclidean" and "l2" possibilities (e.g. pairwise distances, etc.).