Best way to emit/output internal network state (for debugging)

[nathanmartz]

As our networks have grown more complex, we've needed more ability to debug the models "in vivo" which means having the model output not just the action values but some internal values computed at different points in the network. We can then take these internal values and log them or even use them for debug rendering in the environment.

So far we've hacked this up by adding actions to the network, but that's clearly suboptimal, so we wanted to see if there was a recommended way for having the model output "side channel" values, which we could then use to pass out the results of internal calculations.

Any advice or pointers would be much appreciated.