Memcnn

------------------
* Added support for mixed-precision training using torch.cuda.amp (inputs fixed to float32 for now)
* Added support for PyTorch v1.7
* Dropped support for PyTorch < v1.0 and Python 2
* Removed the version limit for Pillow in the requirements

------------------
* Added support for splitting on arbitrary dimensions to the Couplings. Big thanks to ClashLuke for the PR
* Added a preserve_rng_state option to the InvertibleModuleWrapper

------------------
* Improved InvertibleModuleWrapper
* Added support for multi input/output invertible operations! Big thanks to Christian Etmann for the PR
* Improved the is_invertible_module test
* Added multi input/output checks
* Fixed random seed per default
* Additional warning checks have been added

------------------
* HOTFIX InvertibleCheckpointFunction uses ref_count for inputs as well to avoid memory spikes

------------------

* Updated underlying mechanics for the InvertibleModuleWrapper
* Hooks have been replaced by a torch.autograd.Function called InvertibleCheckpointFunction
* Identity functions are now supported
* Reported unstable memory behavior should be fixed now when using the InvertibleModuleWrapper!
* Minor changes to test suite

------------------

* Added InvertibleModuleWrapper support to is_invertible_module test