fixed contextual attention module (various problems & discrepancies with baseline implementation)

[ShnitzelKiller]

• Fixed extract_patches() to permute the dimensions so that the subsequent views are correct (Issue #2)
• Fixed mask patch processing, which was inconsistent with image patches
• Fixed normalization of convolution weights
• Fixed downsampling with non-unity rate
• Added runnable test in model_module.py to test the attention module