Speed up writing mates

[IanSudbery]

Fixes #539.

When TwoPassPairWriter reaches the end of a contig, it calls write_mates, which reopens that contig from the start and scans through for remaining mates. To do this, write_mates uses pysam.AlignmentFile.fetch(...., multiple_iterators=True). As the file uses is the same filehandle being used by bundle_iterator, then multiple_iterators=True ensures that the position in the file is not lost in this operation.

multiple_iterators=True imposes some overhead. With a reasonable number of contigs this is not a problem, as the overhead is small compared to the cost of the scan. However, when alignment is done to the transcriptome, write_mates is called 100s of thousands of times, and for some reason, fetch calls the __init__ of psyam.RowIteratorRegion 4 times for each call to fetch. This causes a serious slow down, such that adding --paired to the commandline slows the processing for an example file down from a couple of minutes to five hours.

This PR changes TwoPassPairWriter so that it's __init__ opens a second file handle to the input file. This allows it to drop the requirement to use multiple_iterators=True and returns the performance to near that of the performance without --paired.

As far as I can tell, this does not change the output (i.e. the two handles act independently). I have tested this both on the test files and on an example transcriptome alignment provided in #539.

Time to run is reduced from 5 hours to 200 seconds.

[TomSmithCGAT]

Wow, that's a pathologically bad performance UMI-tools has had for transcriptome alignments 😱 Good catch!