Inplace-abn

This release updates `ABN`, `InPlaceABN` and `InPlaceABNSync` to feature parity with recent versions of Pytorch's BatchNormNd layers:
* Add a `track_running_stats` parameter to enable / disable computation of running statistics independently from the layer's training state
* Add a `num_batches_tracked` buffer, and allow passing `momentum=None` to support cumulative moving average for tracking running stats instead of exponential moving average
* As a side-effect, now support loading parameters from standard BatchNorm without work-arounds. Still, if the loaded parameters contain negative weight elements the output will differ compared to standard BatchNorm

Additional changes:
* Fix backward pass in eval mode: it was not properly accounting for the activation function
* Refactor code to follow more sensible formatting standards
* Add type annotations
* Improve docstrings
* Update installation instructions, pointing to the PyPI package

This release contains an improved implementation of the fix for the backward pass in v1.0.9 which uses less temporary memory at no additional computational cost.

In previous versions, both the input/output tensor `y` and the gradient tensor `dy` were overwritten during the backward pass. This was causing issues with some network topologies, producing wrong gradients.

To fix this issue, a pair of temporary tensors is now created during the backward pass to hold the results of intermediate computations. This change will increase the amount of temporary memory required, meaning that in some cases where GPU memory utilization was already very close to the limit OOM errors might now occur. An alternative, more complex fix is also possible at the expense of additional computational costs. We are evaluating the impact of these changes and will provide updates in a future release.