Handle NaNs properly from the web UI.

[tshead2]

Currently, we assume that there are no NaNs in the input or output columns when computing a CCA model, and our push scripts / CCA UI enforce this behavior. NaNs in ignored columns are fine, so it is possible to import data containing NaNs, if they're in ignored columns. However, this means that if a user re-runs a model and uses a column containing NaNs, the computation will fail.

We really ought to be filtering NaNs out of the data when we compute the model, so it doesn't matter whether the uploaded data contains them or not. However, this raises an issue with the final visualization, in that we can no longer assume a one-to-one mapping between rows in the raw data and points in the scatterplot.