ReadChooser was storing all the sequences and score in a map (log-time access)
and then reads were stored in a vector which was sorted depending on the
scores stored in the map. So we had many allocations (map+vector) and a slow
access (log access time, so the sort was in n log²n, where n is the number of
sequences). I'm a bit ashamed to have done such an inefficient thing.

Now the scores and the sequences are stored in an array. The sequences
are not copied, but just a pointer to them is stored.

The C function qsort is used because the C++ sort doesn't work
(because of implicit destruction of something).

The code is now much quicker (a -c clones of 600k reads, with -z 1 is now twice as fast).

Someone motivated could probably do better with a radix sort.