Fix single-precision soft equivalence for plane

[sethrj]

This adds documentation to and fixes the soft equivalence function for planes. Due to the comparisons using $\epsilon^2$, a tolerance of 1e-4f can get washed out in the machine precision error which is larger than 1e-8f.

I have also uninlined LocalSurfaceInserter::operator() to reduce code size.

This also fixes the extra semicolon in ConvexSurfaceBuilder that @stognini reported, and the missing -Werror that allowed that to pass the CI.