Better logging

[riley-mld]

At the moment we're missing some important metadata in our logs. Some examples of these are metrics like: buffer size, bp_per_sec, inf_per_sec, bp_per_inf and some info that could be useful for training such as episode reward, episode length. We also miss some other metadata that could perhaps be useful for debugging like weights, biases.
I think replicating what we have in Cog seems like a good place to start.