Decision transformer

[rafiberlin]

I have tested the Decision Transformer in VLN-CE.

While the results were underwhelming, it might still be interesting to integrate this into the main code base. It shows a working implementation with Transformer based agents (how the handling of the multiple environments needs to be done for example).

There is one improvement for run.py. If a directory is given as a parameter, it will run all training files included in the directory. The resulting training log file has the same name as the training file with a .log extension.