Need a better way to control log message verbosity than 'verbose' parameters

[mcloughlin2]

AMPL has a parameter 'verbose' that mainly only gets used when running predictions from a model. Among other things, it sets the logLevel for the ATOM logger to CRITICAL when verbose is False and DEBUG when it's True. In 1.6.1, it leaves logging in that state after the prediction is done, so if you run a prediction with verbose=False and subsequently train a model, the informational messages from model training aren't shown. One of the hackathon students noted this, but it's something that's bugged me for a long time.

In 1.7.0 I changed the prediction functions so that the logLevel is restored before the function exits; that solves part of the problem. However, I still don't like the all-or-nothing behavior of 'verbose' parameters, plus the fact that they have no effect on model training. We need to have a simple function to set the logLevel for AMPL messages. In addition, it would be good if we could control the verbosity level of messages from component libraries like RDKit, TensorFlow, Mordred in a more fine grained fashion.