Fix negative log likelihood loss and its derivative and add unit tests

[Hydrotoast]

Looks like I was missing a negative sign on the derivative for nlogsig_deriv. Integrating this should give us nlogsig, the negative log likelihood.

Should be noted somewhere that labels are intended to be of the form {-1, 1} for correctness. Otherwise, we can transform the labels when the classification task is chosen.

[coveralls]

Coverage increased (+0.9%) to 97.414% when pulling 770fb27 on Hydrotoast:master into 1ea355a on btwardow:master.

[btwardow]

lgtm