Fix CER/WER always equal to 1.0

Closes #24 (closed)
As in the predict.py code (see https://gitlab.com/teklia/atr/dan/-/blob/main/dan/predict.py#L153), to check if the end_token has been predicted, we should use == instead of is since it doesn't work correctly on tensors.

dan/manager/training.py

@@ -1262,7 +1262,7 @@ class Manager(OCRManager):
                 )
                 predicted_tokens_len += 1
 
-                prediction_len[reached_end is False] = i + 1
+                prediction_len[reached_end == False] = i + 1  # noqa E712
                 if torch.all(reached_end):
                     break