Drop fixed-benchmark

This would close #346 (closed).

As advised in that issue, the fixed-benchmark has been deleted. Replacing it with a good benchmark would be difficult, because in actual use wl_fixed_to_double gets inlined by compilers; there is also negligible benefit in optimizing the function, since wl_fixed_to_double is not called on a large volume of data.

Minor note: Two almost equivalent alternatives to the chosen f * (1. / 256.) were f * 0.00390625 (too much magic constant) and f / 256. (not clearly representative of the compiled code due to div->mul optimization at -O1.)