[MRG] Avoid uncessary copies in sklearn.preprocessing

[rth]

Partially addresses #13986

This removes the copy=True in the fit method of StandardScaler, MinMaxScaler, MaxAbsScaler, RobustScaler where it is typically not necessary to compute the scaling factors.

In practice, this makes StandardScaler().fit_transform 10%-20% faster on the few examples I have tried.

If that copy was necessary and this mistakenly removed it check_transformer_general(.., readonly_memmap=True) would fail in common tests.

[NicolasHug]

LGTM.

Looks like in these specific cases inplace would have been a more descriptive parameter name than copy, and might have prevented this.

[thomasjpfan]

LGTM

[thomasjpfan]

Does this need a whats_new entry as an enhancement or a bug fix?

[rth]

Thanks for the reviews! Added a what's new.

[thomasjpfan]

QuantileTransformer has a _check_inputs that copies during fit and transform. What you think about adding copy parameter to _check_inputs and setting it to false during fit and self.copy during transform?

[thomasjpfan]

Thank you @rth!