lfda not deterministic

[wdevazelhes]

See comment #193 (review)
It seems that LFDA is not deterministic because of an eigenvalue decomposition that can return one sign or another as output
More details to come

[bellet]

I think it is fine, this happens also in sklearn, see for instance the "Notes" section:
https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.TruncatedSVD.html

What we could do however is to have a random_state in init (related PR: #171) so that we two instances of LFDA with the same random state and fit once on the same data will give the same result