allow user to modify model prediction results

[mikoontz]

This PR addresses some of the need that arose in #10. In some use cases (not sure how common!), a user may need to modify the truth, response, or prob values output generated by the predict_learner() function. This patch introduces two new arguments to the cpi() function (which get passed to the compute_loss() function): modify_trp and ....

modify_trp ("trp" stands for truth/response/prob) is by default FALSE (leave the truth/response/prob output alone) but can also take a user-defined function that accepts as arguments the original truth, response, and prob values (after the modification that takes place in the beginning of the compute_loss() function) as well as the ... argument to allow for some flexibility. It must return a named list (names must be truth, response, and prob) with the updated values. If values don't need updating, a user can simply pass the original value to the corresponding list item in the returned object.

Some potential use cases (which I've added examples for in the docs):

1. A user wants to use a different classification threshold than 0.5 for choosing which level to classify a particular observation
2. A user is building a random forest regression on a 0/1 binary outcome in order to use an unbiased split selection criteria (e.g., the maxstat rule, which requires a regression approach not classification; further discussion here). That user then still wants to use classification-like loss measures and therefore needs to create a prob object from the response object, and to update the response object to represent the predicted class.
3. A user wants to rescale the probability output (as in consequences of logloss as loss measure for binary classification when optimal classification threshold != 0.5 #10)

It strikes me that I can imagine the modify_trp and ... arguments being passed to the predict_learner() function instead of compute_loss(), but the machinery at the start of compute_loss() makes the truth, response, and prob objects a little easier to work with (because they're made consistent regardless of whether inherits(pred, "Prediction") is TRUE or FALSE). But this block could also move to the end of predict_learner() to make the pred output consistent at this stage?

if (inherits(pred, "Prediction")) {
    truth <- pred$truth
    response <- pred$response
    prob <- pred$prob
  } else {
    truth <- do.call(c, lapply(pred, function(x) x$truth))
    response <- do.call(c, lapply(pred, function(x) x$response))
    prob <- do.call(rbind, lapply(pred, function(x) x$prob))
  }

Just a thought on organization (keeping all the prediction modification within predict_learner() instead of spread across predict_learner() and compute_loss()) to discard if not useful!

I think this approach allows maximum flexibility for a user, but I'm also happy to contribute to further discussion on other ways to implement something like this (if it might be implemented at all).

All checks pass and I added some worked examples to the documentation.

Thanks again for the great package!