LSML fails on some classical inputs

[wdevazelhes]

while finalizing #180, I encountered problems with LSML: examples that shouldn't be problems fail: here are a snippets that failed

import numpy as np
from sklearn.datasets import make_classification
from metric_learn import LSML_Supervised

X, Y = make_classification(n_classes=3, n_clusters_per_class=2,
                           n_informative=3, random_state=42, class_sep=4.,
                           n_features=5)

lsml = LSML_Supervised(num_constraints=200, tol=0.0001,
                       max_iter=10000, verbose=True)
X_lsml = lsml.fit_transform(X, Y)

The error returned is: ValueError: The input metric should be symmetric.

[wdevazelhes]

Turns out the problem was that in this example there are two redundant features i.e. that are a linear combination of other features (n_redundant=2 by default) and therefore by default for the prior LSML will compute the inverse of a covariance matrix that has two null eigenvalues (at this line: https://github.com/metric-learn/metric-learn/blob/9f732502dbf778c8c579cffc3927f619ed8f4089/metric_learn/lsml.py#L73)
Replacing np.linalg.inv by scipy.linalg.pinvh (the pseudo inverse) solves the pb
I'll open a PR for that (but anyway this will be solved also in #195), it's just that this will be easier to merge and will allow to finalize #180