After having tried to deal with a few issues related to extending Imputer behaviour, I believe we should be removing the axis parameter from Imputer.

• It seems a strange feature to support in a machine learning context, except perhaps where the features represent something like a time series.
• It is not stateful and can be performed with a FunctionTransformer. (We could even provide a row_impute function, if we felt it necessary, which would roughly be defined as def row_impute(X, **kwargs): return Imputer(**kwargs).fit_transform(X.T).T.)
• It complicates the implementation, which already has a bunch of weird edge-cases (to handle sparse data with missing indicated by 0 which is an inefficient use of a sparse data structure; and to handle non-NaN missingness indicators), unnecessarily.
• It is often nonsensical to extend further features to the axis=1 case.

Do others agree?