Improve accuracy and safety of float serialization

Multiplying by a factor of 1/1000.f (rather than dividing by 1000.f) directly
introduces an error of 1 ULP.  With this patch, an exact comparison of a
floating point literal with the deserialized F1000 form representing it is now
guaranteed to be successful.
In addition, the maxmium and minimum safely representible floating point
numbers are now well-defined as constants.