Note [Rules for floating-point comparisons]

We need different rules for floating-point values because for floats
it is not true that x = x (for NaNs); so we do not want the equal_rule
rule that mkRelOpRule uses.

Note also that, in the case of equality/inequality, we do /not/
want to switch to a case-expression.  For example, we do not want
to convert
   case (eqFloat# x 3.8#) of
     True -> this
     False -> that
to
  case x of
    3.8#::Float# -> this
    _            -> that
See #9238.  Reason: comparing floating-point values for equality
delicate, and we don't want to implement that delicacy in the code for
case expressions.  So we make it an invariant of Core that a case
expression never scrutinises a Float# or Double#.

This transformation is what the litEq rule does;
see Note [The litEq rule: converting equality to case].
So we /refrain/ from using litEq for mkFloatingRelOpRule.