investigate bottom layer in TernausNet implementation

[taraspiotr]

In TernausNet at the bottom:
https://github.com/neptune-ml/open-solution-mapping-challenge/blob/865f37571a844d698ac9ac898eb02249c9d6c796/src/unet_models.py#L392
we are pooling already downsample layer, we should investigate if:

• we can remove this layer and have the same accuracy with less parameters
• we improve accuracy with changing pooling into conv layer with stride 2