Improving masking mechanisms in CNN and LightAttention models

[SebieF]

• Applying masking to CNN: Padded residue embeddings are now masked out which improves reproducibility and avoids different predictions between batches and single inputs
• Improving masking in LightAttention model: Mask is now applied before the attention convolution, and -float('inf') is used instead of -1e9 which seems to improve reproducibility and avoids different predictions between batches and single inputs
• Adding inference unit test to check if batch predictions match single predictions
• Updating models and related inference unit tests

Closes #100.