make output transform unpad and scale down mask in penobscot metrics - facilitates correctly reported inline mIoU

[maxkazmsft]

Address the following TODO in interpretation/deepseismic_interpretation/penobscot/metrics.py

# TODO: make output transform unpad and scale down mask
# scale up y_pred and remove padding
class InlineMeanIoU(Metric):

In the same class, in update method, add assertion to cover the dimensionality of max_prediction variable

        # TODO: Make assertion exception
        # issue: https://github.com/microsoft/seismic-deeplearning/issues/276
        max_prediction = y_pred.max(1)[1].squeeze()