NaN score checks

[jrmyp]

Franck Buijs (currently in internship) uses an attelo-independent learner for his experiments, and got hard-to-debug behavior upon decoding.

NaN scores passed silently to the MST algorithm until something crashed the algorithm.

This PR adds a check to the MST decoder for NaN scores. I'm investigating if the check should be propagated to the other decoders, or if the check should be made at a higher level.

[moreymat]

The Eisner decoder explicitly handles NaN values in a few places.
These were a quick fix, I did not try to come up with a general solution to cleanly handle NaN values between attelo learners and decoders.

Note however that there is not a unique solution to NaN values, they need to be traced and addressed on a case-by-case basis.
Some NaN values can be avoided if we check for certain parameters in the computation, some cannot and should be intercepted later (ignored or replaced by an appropriate value in a given context).
This might be easier with more information about the code @fbuijs is developing.