[MRG] Move transformer_from_metric to util

[bellet]

Fixes #146

[wdevazelhes]

Seems great !

[perimosocordiae]

LGTM, merged.

[perimosocordiae]

Side note: we may want to find a more performant way to check if a matrix is diagonal, other than allclose(m, diag(diag(m))). I wonder if scikit-learn has something along those lines?

[perimosocordiae]

Then again, the metric is a d by d matrix, so it's not a big deal in most cases.

[wdevazelhes]

Yes, according to this answer and this answer, maybe a good way in numpy would be to remove the diagonal and check if the matrix is zero, like:
np.fill_diag(m, 0); not m.any() (but apparently any does not short circuit when finding a not zero element as noted in the Stackoverflow answer (but then this answer uses numba to do a short-circuit function)
Plus I don't know if there is a way to test if numbers are close to 0 in numpy
And yes maybe scikit-learn has some helper function to do that