In python interpreter make streams unbuffered by default

[astaff]

With buffered streams, when running python scripts users won't be seeing realtime logs in neuro logs. Starting python interpreter as python -u or setting PYTHONUNBUFFERED=1 will make log streams nearly realtime.

I suggest setting PYTHONUNBUFFERED=1 in Dockerfile to make it a default for every python interpreter run assuming that the users are not gonna pay attention and run their training scripts with python -u

[atemate]

is this option required for all jobs? Note, it might slow down training that writes to logs a lot, see aws/amazon-sagemaker-examples#319.

Maybe we should still close this issue on the Template side: neuro-inc/flow-template#264

[mariyadavydova]

Well, I agree with the following assumption:

the users are not gonna pay attention and run their training scripts with python -u

We have a problem with non-real-time logs right now, which this change gonna fix. It might slow down the train process, but we'll see and solve this problem later.