Fix issue_703: backward compatibility with python3.6

[goodok]

Before submitting

• Was this discussed/approved via a Github issue? (no need for typos, doc improvements)
  • Fitting with log_gpu_memory=True fails in python3.6. #703 (there are no comments)
• Did you read the contributor guideline?
• Did you make sure to update the docs?
  • No visible changes
• Did you write any new necessary tests?
  • About py.test -v
    Distributed modes don't work on my machines.
    test_amp.py::test_amp_gpu_ddp freezes.
    It is not related with pytorch-lightning, and it is not related with this PR (master branch freezed too, in pure pytorch too, other ML too).
    So I can't run the tests and I tested them manually (examples without distributed modes)

What does this PR do?

Fixes #703
Backward compatibility with python3.6
Fitting with log_gpu_memory=True in the Trainer fails in python3.6 version.
Reason:
subprocess.run is used with capture_output which is valid only for the python >= 3.7
See also workaround to maintain python3.6:
https://stackoverflow.com/questions/53209127/

PR review

Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

I'll wait.

[Borda]

LGTM, may a test here would be nice...