Implement convac in torch instead of NumPy

[idfah]

Description

The current NumPy-only implementation of convac / CNN-LA is somewhat difficult to maintain because of manual gradient computation. It also does not support GPUs. This story is to re-implement this network architecture in pytorch with optional GPU training support.

Acceptance Criteria

• New implementation of convac available and written in pytorch
• New convac implementation supports both CPU and GPU training

Tasks

• create new module structure for torch re-implementations
• implement convac
• add GPU support
• add unit tests