looking at the linked post. The difference in Winmark is tiny. Can it be repeated? Some subtests seem to be quicker without FPU which suggest the benchmark run wasn't very stable. I'd consider those results as a fluke.
Win G or the benchmark could use some FPU code, I am not sure about that. But this is not a standard part of Windows.
I can't think any way how to speed up Win3.1 driver operation or GUI acceleration using floating-point math in Win 3.1 days. Even if there was you couldn't just release FP code in drivers back in those days as FPU wasn't standard and Windows had to run down to 286. You would need two sets of drivers or some FP emulation linked it which wouldn't be practical for drivers that are only a few kilobytes...
SMARTDRIVE source code is public, I've checked that and there is definitely not any FP code. I doubt there is anything in virtual memory management that would benefit from having a FP.
Now I am motivated to scan my Windows installation and drivers for FP opcodes. I am not convinced...