Gradient-metrics

Fixed `GradientMetricCollector.__init__` which caused `__del__` to raise an `AttributeError` when supplied with wrong parameter combinations.

Also added a test for `__del__` and restructured the tests into more functions for better overview.

First published version of the package.
It supplies utilities for computing gradient metrics on neural networks using PyTorch. Metrics are collected by utilizing backward hooks.

Currently the following metrics are supported:

PNorm
Min
Max
Mean
MeanStd


`GradientMetricCollector` is a helper class for collecting the metrics and computing the gradients.