Performance metrics - report average class accuracy, rather than just accuracy.

As discussed in issue #237 and PR #238 , the training and validation results can be biased by the class imbalance typical in high-throughput experiments. Negative controls tend to have a lot more example images, both for training and validation, while other classes have only a few. The accuracy metric, which is the default in DeepProfiler training and validation, is sensitive to class imbalance. Therefore, the observed performance may not inform correctly how many phenotypes can be recognized, only how many images regardless of the class.

We should use average class accuracy to obtain a metric that gives an estimate of how many classes can be accurately recognized from the images, instead of how many images are correctly classified independently of their class.

As far as I know, Keras does not implement a metric like this. But we could easily implement a metric plugin and make it the default performance metric in DeepProfiler.