implement a new bam filtration step

[dpark01]

Picard FilterSamReads is both slow and doesn't scale well with memory: it appears to store an in-memory map of all read IDs given to it as input. Larger NovaSeq runs seem to require VMs with > 8GB of RAM when running any kind of filtration step. We should make good on this vow in the code comments from long ago about replacing Picard FilterSamReads with some combination of pysam, sqlite, and is O(1) memory with respect to the size of the read ID list (and the BAM file).