Compute single-end MAPQ score better by considering at least two mappings.

[ksahlin]

We knew for a long time that our MAPQ assignments have not always been the most representative. We should consider whether we want to make such an effort (see also the PE mapq fix in PR #596).

This commit prevents most (but not all) of the false MAPQ=60 assignments. In some cases, we also get a couple of more correct alignments (although very few).

It does slow down the mapping a tad, I did some preliminary benchmark below based on some datasets I had generated for other analyses (diagnosing the sweep PR). The outlier of -18% in runtime is just an anomaly because I benchmarked on my laptop. Overall, it seems we fix 40–98% of false MAPQ=60 assignments, at a cost of <2% increease in runtime.

However, I also looked at 10,000 bp reads, and the runtime increased by 6%, while we corrected nearly none of the (much fewer) false mapq=60 assignments. So this is an example where we may not want to do the same for long reads, as discussed here: #599

Corrected MAPQ

genome	length	tag	correct	wrong	false_mapq60	unmapped
chrY	100bp sim0	main	20,237	29,727	416	36
chrY	100bp sim0	main_fix	20,320	29,644	183	36
chrY	250bp sim0	main	24,089	25,908	526	3
chrY	250bp sim0	main_fix	24,023	25,974	13	3
chrY	1000bp sim0	main	34,016	15,983	223	1
chrY	1000bp sim0	main_fix	34,066	15,933	1	1
CHM13	100bp	main	457,556	41,801	798	643
CHM13	100bp	main_fix	457,588	41,769	460	643
CHM13	1000bp	main	484,434	15,559	170	7
CHM13	1000bp	main_fix	484,431	15,562	83	7
maize	100bp	main	36,010	13,958	134	32
maize	100bp	main_fix	36,025	13,943	25	32
maize	250bp	main	45,258	4,742	44	0
maize	250bp	main_fix	45,255	4,745	2	0

Runtime

genome	length	main	main_fix	delta
chrY	100bp	3,540 ms	3,558 ms	+0.5%
chrY	250bp	4,528 ms	4,504 ms	−0.5%
chrY	1000bp	15,553 ms	15,839 ms	+1.8%
CHM13	100bp	15,008 ms	12,334 ms	−18%
CHM13	1000bp	39,287 ms	40,152 ms	+2.2%
maize	100bp	53,884 ms	54,385 ms	+0.9%
maize	250bp	55,083 ms	55,218 ms	+0.2%