allow for custom loss function to be passed as measure argument

[mikoontz]

Allows for a custom loss function to be supplied to the measure argument. Must take truth, response, and prob as arguments (as they are output from the predict_learner() function followed by the few modifications at the beginning of the compute_loss() function (which themselves might be further modified by a call to the modify_trp() function). Returns a single vector (which may be a one-element vector) representing the loss (i.e., 1 - model skill measure).

Addresses some of the need described in #10.

I added an example to the help file that shows an implementation of Matthew's correlation coefficient. I included a note in that example about MCC being a batch-level skill measure and so that statistical approach currently implemented in this package isn't yet set up to handle that. The work around is to have the user set up a series of test_data data frames ahead of time, which may mimic a resampling strategy (like repeated cross-validation, with some number of iterations and folds defined), then to use that series of batch-level CPI values to conduct an external significance test.

All build tests and checks pass. DESCRIPTION and cpi.Rd files changed using {roxygen2}.

[mnwright]

Thanks! I might want to simplify the example a little bit but I think this is very useful.

I'll als check with mlr3 if we could have some way to use all their implemented measures, i.e. without averaging.