Only training alignment model for marker placement on translated content

[Enkidu93]

This issue also exists in machine.py/Serval: sillsdev/machine.py#320. See there for more details.

This is slightly more complicated to fix here in silnlp since we'll have to supply training data to the postprocessor (or train the alignment model during training and reference it somehow in postprocessing) whereas in machine.py, we have access to the training and inference data in one place already.