Add checks that number of matrix elements do not exceed max int in CreateReadCountPanelOfNormals.

[samuelklee]

Ran into this when trying to create a PoN with 100 samples x 100 bp bins = 1.27 * max int elements.

This currently causes issues when truncating outliers, at which point all elements are loaded into an array so that Percentiles can be naively computed, resulting in a java.lang.NegativeArraySizeException.

Solutions include: 1) simply throwing a message and failing early if the counts matrix is too large (perhaps recommend scattering by contig, see #4728), 2) changing the outlier truncation procedure to be more robust.

I'm not sure how important outlier truncation is to the SVD, as it remains to be evaluated, but for now we should be able to get around this with no code changes by simply disabling it (i.e., setting the relevant truncation percentile to 0).

Note that file I/O takes about an hour for this case. Also note that this is probably on the extreme end of what we should expect to support on a single machine with all counts in memory, as the SVD is probably sufficiently good with 100 samples and 100 bp is on the order of the read length. #4728 will get around this and also make downstream tasks complete faster in parallel, at the very small expense of reducing a few global parameters to per-contig parameters in the modeling step.