Starting from the full read, we can not limit the DP computations
to a k-band around the diagonal without first knowing where is exactly the junction.

Nevertheless, we can avoid computing about one half of the DP matrix,
as the end of the V / the start of the J (minus some deletions) must be matched.

The BOTTOM_TRIANGLE_SHIFT is now set to 20, and this should be large enough
to handle V/J deletions until ~30 bp (see comment in segment.h).
(The current tests were even passing with BOTTOM_TRIANGLE_SHIFT set to 10.)

Now the FineSegmenter (as launched by 'make shouldvdj_with_rc_merged') is about 35% faster.