Wrap train-taskcluster.sh in train_taskcluster.py

[bhearsum]

In order to support automatic continuation of spot terminated training runs (#270) I'll need to do some non-trivial things before handing off to train.sh. Rather than try to do them in bash, I'd like to do them in Python. This PR should have no behavioru differences, but it lays a foundation for this upcoming work. It would also be a good basis for replacing train-taskcluster.sh with Python entirely in the future.

[eu9ene]

I'm concerned about the number of wrappings in our training. We have: Marian wrapped by OpusTrainer wrapped by train.sh wrapped by train-taskcluster.sh that will be wrapped by train_taskcluster.py...

I strongly recommend rewriting train-taskcluster.sh in python instead of adding another wrapping level to reduce complexity.

[bhearsum]

I'm concerned about the number of wrappings in our training. We have: Marian wrapped by OpusTrainer wrapped by train.sh wrapped by train-taskcluster.sh that will be wrapped by train_taskcluster.py...

I strongly recommend rewriting train-taskcluster.sh in python instead of adding another wrapping level to reduce complexity.

I'm a bit hesistant to do take that on as something that blocks turning spot instances back on, since we're on a tight timeline, and it's difficult to test all of the training continuation scenarios, but I can if you'd like.

Alternatively, I can follow up with it later, after we've branched for the upcoming trainings.

[eu9ene]

I'm concerned about the number of wrappings in our training. We have: Marian wrapped by OpusTrainer wrapped by train.sh wrapped by train-taskcluster.sh that will be wrapped by train_taskcluster.py...
I strongly recommend rewriting train-taskcluster.sh in python instead of adding another wrapping level to reduce complexity.

I'm a bit hesistant to do take that on as something that blocks turning spot instances back on, since we're on a tight timeline, and it's difficult to test all of the training continuation scenarios, but I can if you'd like.

Alternatively, I can follow up with it later, after we've branched for the upcoming trainings.

Sure, it's up to you, fixing this as a follow-up works for me.

[eu9ene]

I'm ok with merging this to unblock spot instances but please add a TODO and an issue to rewrite this in Python to reduce wrapping and complexity.