Removing the default requirement of python created by the virtualenv.
This is due to updates on the remote servers, so 3.8 is no longer
default. Note that this may (and has) caused bugs with
from_checkpoint.py - All RLlib algorithm checkoints are saved using
pickle, which is python-version-dependent !!!

Speaking of from_checkpoint.py, it currently works for PG/PPO
non-recurrent policies only ! QMIX is broken (again :(), and the
recurrent policies need more arguments when calling
compute_single_action.