Cannot choose initial metric in the same way for Mahalanobis metric learning algorithms

[bellet]

Depending on the Mahalanobis metric learning algorithm, the possible choices for the initial metric are not the same. For instance MMC and ITML accept an initial metric as parameter, NCA has none. The default initial metric also varies from an implementation to the other: for instance MKLR initializes by default with PCA.

There could be some unified interface to set the initial metric for Mahalanobis algorithms