rewriting f_oneway in the scikit to avoid useless recomputations

and 3 duplicates of the data in memory. It should probably be
fixed directly in scipy. Adding test to make sure that both
functions give the same results + a bit of PEP8