Review default LR scheduling parameters for training

[kprokofi]

Current values for MaskRCNN looks too strict, making training on difficult datasets suboptimal Related reported issue: https://github.com/open-edge-platform/geti/issues/1603

Consider changing them and check other models as well