demultiplex by lane, then merge

Processing NovaSeq flowcells is very slow, due to the bottleneck of demultiplexing, which is currently linear for the whole flowcell. This can be improved by demultiplexing each lane separately, then merging the resulting bams. 