Merge returnn_training and returnn_rasr_training?

[christophmluscher]

This function shares almost the same signature as returnn_training and does very similar things.
Would it be possible / desirable to only have one function, which then internally decides on ReturnnTrainingJob vs ReturnnRasrTrainingJob and does the handling based on *_corpus_key.

Originally posted by @michelwi in #30 (comment)