[ENH] subsampling in Clustered Inference

[jpaillard]

The subsampling used in CluDL currently uses sampling without replacement. Should we add an argument bootstrap: bool to allow sampling with replacement?
I think the problem with the current situation is that using a sampling fraction of 1 leads to unexpected behavior, where the same analysis is run repeatedly when n_bootstraps > 1 in EnCluDL.

[bthirion]

IIRC the sampling is just used for the clustering part ?
We could simply impose sampling with replacement, but sampling without replacement makes the clustering deterministic, which can be useful for debugging. So adding bootstrap=True makes sense indeed.