Include variance in NN policy for ros agent

[cbfinn]

This PR adds functionality to run neural net policies with non-zero variance on the PR2/agent_ros. This is crucial for running MDGPS with agent_ros.

This PR also adds a 2 second wait after the agent_ros reset, which is important for running experiments on the real robot, but not necessary for gazebo experiments (and will slow down gazebo experiments).

Lastly, it includes a minor bug fix in neuralnetworkcaffe.cpp.

[cbfinn]

@anuragajay Can you double-check that I got all of the changes in this PR?

[anuragajay]

Yes. It includes all the changes.

[cbfinn]

Thanks!