Unexpected run structure when running Acme DQN agent

[nickjalbert]

Copied from PR 227 Review (in a comment hidden by the dumb GH comment accordion), confirmed this happens to me in master:

I'm running into some sort of issue with acme_dqn agent (probably both the acme agents?). Repro:

cd example_agents/acme_dqn
rm -rf mlruns/
agentos run agent --param-file parameters.yaml --entry-point learn

I get the following (unexpected) run structure:

In the learning child run, I see metrics and parameters but no artifacts. In the free floating run, I see the model artifact but no metrics and parameters. Probably something with the way we terminate and clean up the run?

Additionally, when I then run:

agentos run agent --param-file parameters.yaml --entry-point evaluate

It says it cannot find the saved model from training.