Correct pnegate for floating-point zero.

The original formulation of 0 - x incorrectly handles +/-0 for floating point numbers. We need to instead flip the sign bit explicitly.