Fix assertion failure in dfdx and dfdy lost bit calculations.

Fix assertion failure in dfdx and dfdy lost bit calculations.

- Fix lost-bits calculation returning negative values when operand
  component range intercepted zero (i.e. operands had different signs).
  This caused verification to expect more significant bits than the
  native (lowp/mediump/highp) float format could store. With highp
  values, error calculation tried to create a error compontent of a
  floating point value with 24 mantissa bits which cannot be stored
  to a native float32.
- Use better terminology, s/nominator/numerator/.

Bug: 19694789
Change-Id: Ifa88fb1c8cb333bab22a550d01f5e310d2d2673d