Fix: random seed set to int if not provided, support reproducible runs

[amoodie]

set random seed to an int in 0 to (2**32)-1 if it is not specified in the YAML configuration file. This enables runs to be reproducible, even if the seed is not set explicitly.

closes #58

[ericbarefoot]

I see, so we set the seed by default and record it, which gets around the whole "random states are easily recorded and reinstantiated." problem.

However, I am curious how we should handle starting from checkpoints in this framework. Should we also eventually include a way to record the random state and just start by using the random state?

[amoodie]

I am curious how we should handle starting from checkpoints in this framework. Should we also eventually include a way to record the random state and just start by using the random state?

Good question, I hadn't thought about checkpointing.

I think the way to implement checkpointing (#57) is to write a single .npy file to disk every N timesteps. In the OG checkpointing implementation, the grids are all written to different .npy files, but it should be straightforward to pack the grids, time-status, etc, into one file and write to disk.

So, for the random state, I agree with your idea. We would do a quick rs = np.random.get_state(), and then pack rs into the checkpointed .npy file. Then when loading, just set the random state to rs.