This avoids vectorizing SLP subgraphs that just compute uniform
operations on all-same operands. That fixes the less interesting
(but most embarrasing) part of the testcase in the PR. On the
way it also fixed a missing matches[0] reset in the last
refactoring touching that place.