NaN handling

[philippberens]

We should decide how to handle NaNs in submission (or how many to allow) and implement it

[freeman-lab]

The current implementation now removes any rows that have a nan in either the measured or predicted rates prior to downsampling. In addition, when running the command line tool to score an entire dataset, we take the nanmedian of scores across neurons. In my testing I occasionally found NaNs in the output of the optimization function for the loglik computation for just one neuron, though I wasn't entirely sure why, probably worth tracking down.