Fix inconsistent marking with coordinate sorted bams in ReadsSparkSource

[jamesemery]

Currently there is a bug in MarkDuplicatesSpark where for non queryname sorted bams the index is incorrectly assumed to be sufficient for reads groups, which wis guaranteed to be adequate for queryname sorted bams. This means that under some circumstances reads can be inconsistently marked as duplicates when its mate is a non-duplicate if they were spread across multiple partitions.

There is a disabled MarkDuplicatesSparkUtilsUnitTests test in #4656 that should hopefully demonstrate this issue.