The generated function, bsearch-fixnum/index, is much faster
because it declares the type of the element inside the cvector.

SBCL took 40 seconds to finish the RUN-CVECTOR-SET benchmark
using standard binary-search. This went down to 2 seconts using
the macro-generated bsearch-fixnum/index.

I do not recall when exactly this began -- but I do remember
that in early versions of Spartns, we did get the same timing
(~2s) in this benchmark. So in some point in time, I did seriously
brreak Spartns' performance, but it's fixed now.