Do not normalize tensor for attention map visualization

The visualization of the attention map looks strange because the image is normalized.

• This command:

teklia-dan predict --image left.jpg --model dan_simara_x2a_left/checkpoints/best_625.pt --parameters dan_simara_x2a_left/parameters.yml --charset dan_simara_x2a_left/charset.pkl --attention-map --output dan_simara_x2a_left/attention/

• Gives the following visualization (right):

Without normalization (old behavior)	With normalization (current behavior)

• We need to change the input tensor in plot_attention to a padded, preprocessed, but non-normalized tensor.