Correction: it looks like two bugs to me.

I think the wrong-return-type bug is rather serious, possibly a release 
blocker for 3.0.1.  I'll see if I can produce a patch.

The incorrect rounding (returning 30.0 instead of 20.0) is less serious, 
but still needs fixing.  For a start, we should eliminate the use of 
pow() in float_round, and when the second argument to round is negative 
there should be a division by a power of 10 rather than a multiplication 
by the reciprocal of a power of 10.