MAJOR decrease in RAM usage.\
With this PR the support for spatially separated convolutions will be removed, however, support for depthwise separable convolutions will stay.
Before merge, the following things should be implemented:
[ ] Reversible Attention
[ ] Reversible Bottleneck Block
[ ] Depthwise separable reversible convolutions
[ ] Removal of Inception (as XCeption/Depthwise seperable convolutions are an eXtreme versrion of it)
MAJOR decrease in RAM usage.\ With this PR the support for spatially separated convolutions will be removed, however, support for depthwise separable convolutions will stay. Before merge, the following things should be implemented: