Construct "masked" zebrafish matrix for use as gold standard for imputation challenge

To identify new and/or improved methods for imputed missing values in our species compendia, we need to put together data that allows us to evaluate performance. This requires at least two matrices: 1) the complete matrix that contains the true values and 2) a matrix where some of the values have been replaced with NAs ("masked").

We want all missing values to have a _true_ value associated with them -- i.e., no values in the complete matrix should be missing. This very likely means subsetting to genes that are only on the `zebrafish` Affymetrix microarray.

Here are the features of the masked matrix that we want:
- Some values to be missing completely at random (MCAR)
- Some values that are missing for some non-negligible number of samples ("missing rows at random"; see #5)
- Some values missing for all microarray samples -- this represents the real challenge of measuring genes in RNA-seq that were not on the chips for the legacy data
- Nice to have: masked values in RNA-seq data that reflect what we've observed -- shorter genes and genes with low expression values are more likely to be zero