The MUnbox instructions the transpiler inserts for GuardIsNumber break the float32 optimization. Fixing that improves performance quite a lot on the f32 assorted tests and also lets us support the assertFloat32/MAssertFloat32 check.

I think we should leave this for now. The float32 analysis is complicated and I'm not confident about the patches I wrote for this.

I want to spend some more time on this after Warp and either simplify or remove the analysis. Maybe we can disentangle it from type analysis and make it a stand-alone optimization pass that still handles the simple (maybe just non-phi?) cases.