a more robust estimation of minimizers repartition. before, short reads could lead to incorrect estimation as we only read a fixed number for reads;
but now, we require to read a certain number of superkmers. will see if that fixes the OAHash max rehashes bug.

still takes quite a bit of construction memory though; also, slightly longer query times. could be solved by switching to compute_mphf_seq but i wanted to try with superspeedy construction.

 related to  why we can't just ignore non-branching nodes forever during simplifications (they may become dead-ends after tip removal);
this removes the possibility of a faster nodes iteration (that wouldn't compute a mphf index for each node)