Get rid of one_minus_pvalues

[bthirion]

One_minus_pvalue is used in DL, CluDL and EnCluDL to denote strong effects with negative sign. I think that this is not rigorous.
What we are actually doing in that case is an unsigned test, that gets a p-values, which needs to be kept, and that's all.
Concretely, one need to

• add a two_tailed_test=True keyword to make it explicit if we want only positive effects or not
• when displaying the significance map in the two-sided test, recover the sign information (which is nothing but the size of the average coefficient at the position) for plotting differently positive and negative significant regions.
  That would be more standard hence cleaner.

[jpaillard]

My suggestion would be to remove one_minus_pvalues, and add the two_tailed_test argument to the fdr_selection class method.
However, since it only makes sense for DL, CluD, L, and EnCluDL, I would not include this argument for other methods, or at least set it to False by default in BaseVariableImportance.
Should we overwrite the fdr_selection class method for DL, CluDL, and EnCluDL in order to allow two-tailed testing? Or put the argument by default to False in BaseVariableImportance and let the user change it when using DL ... ? Or do you have a different suggestion?

[bthirion]

But why do you associate it with fdr_selection ? When we use CluDL and EncluDL the fdr becomes hard to interpret because of the spatial relaxation. i would rather push for FWER control with these classes.

[jpaillard]

Should we raise a not-implemented error for fdr_selection with a message suggesting the use of fwer_selection instead, and implement the latter?

[bthirion]

No, it's not error, but we should make that the docstrings, doc and examples nudge users toward fwer control.