[DOC] Improve documentation based on feedback

[thomasbazeille]

#36 Underline the ability to pass a non-default estimator in PairwiseAlignment(method=)
#37 Insist on the need of several sample to learn a robust alignment
#37 Maybe better describe the ideal use of PairwiseAlignment, with a masker, automatic clustering
#34 Point out that for now, Optimal Transport should only be used on regions with less than 1000 voxels and include profiling figure somewhere