Fix handle_missing parameters and standardize input data shape for MinHashEncoder

[alexis-cvetkov]

I have made small modifications to the MinHashEncoder and GapEncoder:

• In MinHashEncoder, the default parameterhandle_missing="" becomes zero_impute, since we actually do not impute missing values with an empty string. Instead we assign them a encoding vector filled with zeros.
• Since other encoders take as input data with shape (N_samples, 1) (to be consistent with scikit-learn), I updated the MinHashEncoder to behave in the same way.
• In GapEncoder, the default parameterhandle_missing="zero_impute" becomes empty_impute, since we impute NaN with an empty string "".

[GaelVaroquaux]

Can you fix the conflicts and update the changelog please.

[LilianBoulard]

LGTM, thanks for the contributions !

[alexis-cvetkov]

Merging !