Make predict_time_value make sense when recent data is missing in create_train_and_predict_matrices

[brookslogan]

modeltools::create_train_and_predict_matrices currently assumes that the latest time_value in the input data should be predict_time_value. However, this assumption does not appear to be documented, and appears to be violated, e.g., in animalia::production_forecaster, when the latest data available for a signal is not for the forecast date. (Which is all the time for all covidcast signals that I am aware of. Moreso for sources like hhs that can regularly lag by a few days.)