HaplotypeCallerSpark Enable downsampling in the AssemblyRegionWalker

[jamesemery]

After #5416 one of the lingering potential sources of difference between HaplotypeCallerSpark and HaplotypeCaller are in the downsampling, which could cost Spark both correctness and performance at pathological sites. This likely requires #5437 or some equivalent change to be implemented so we can save ourselves from materializing and shuffling all the reads in their AssemblyRegions.