Engine read caching.

[cmnbroad]

Allow tools to opt-in for read caching (enabled in this PR for CNNScoreVariants and FilterAlignmentArtifacts) when issuing forward-only queries such as those used to query a ReadsContext for all reads backing each variant.

The cache matches the (sometimes surprising - see https://github.com/broadinstitute/gatk/issues/4901) results that are returned when caching is not used, specifically that an unmapped but placed read that is mated with a mapped read is only returned if the start position overlaps the interval query interval, whereas the mapped read is returned if any part of the read overlaps the query interval.