Make find_start_sequence use the BNDM search algorithm. We handle

edge cases via the old slow mechanism.

Previously, find_start_sequence would be responsible for ~25% of the
usr CPU time in tests (as performed by Ian).  No more.


git-svn-id: https://svn.apache.org/repos/asf/httpd/httpd/trunk@90940 13f79535-47bb-0310-9956-ffa450edef68