ReadCountCollection pre-processing filters

[mbabadi]

At the moment, we:

• Remove targets with possibly bad (NaN, infinity, negative) values
• Remove targets that have uniformly low coverage across all samples

Perhaps we should consider adding more filters:

• Remove targets with very high and very low GC content (can be done in the CalculateTargetCoverage step)
• Remove targets with lots of repeats and anomalously low mappability (can be done in the CalculateTargetCoverage step)
• In the learning mode, remove a target if too many are masked across the samples (in that case, max likelihood parameter estimation is unreliable)

This must be done after careful evaluations, i.e. only if certain features makes a target prone to bad calls no matter what.

[droazen]

Issue moved to broadinstitute/gatk #2992 via ZenHub