[metrics] Handle metrics updates inside TrainableModel

[generall]

Currently it is required to write metrics update logic manually in the custom implementation of TrainableModel.
But there is a large group of metrics which are updated in similar way, therefore it is possible to relieve the user from it.

Proposition

Implement metrics configuration similar to loss - create an abstract function in the base class configure_metrics.
If re-refined configure_metrics returns non-empty dict, TrainableModel should register metric updates in related hooks

[joein]

Currently, every metric computes distance matrix each time compute method is called, might be reasonable to introduce some kind of cache inside a metric and some kind of metrics collection to share distance matrix between them