Make floating-point exception error messages slightly more verbose: in

Make floating-point exception error messages slightly more verbose: in
particular, the error message now allows one to distinguish between a
ValueError arising from a singularity (e.g. log(0.)), which would
usually produce +-infinity in non-stop mode, and a ValueError resulting
from an invalid input (e.g. sqrt(-1.)), which would normally produce a
NaN in non-stop mode.