Note [Do not eta reduce PAPs]

I considered eta-reducing if the result is a PAP:
   \x. f e1 e2 x  ==>   f e1 e2

This reduces clutter, sometimes a lot. See Note [Do not eta-expand PAPs]
in GHC.Core.Opt.Simplify.Utils, where we are careful not to eta-expand
a PAP.  If eta-expanding is bad, then eta-reducing is good!

Also the code generator likes eta-reduced PAPs; see GHC.CoreToStg.Prep
Note [No eta reduction needed in rhsToBody].

But note that we don't want to eta-reduce
     \x y.  f <expensive> x y
to
     f <expensive>
The former has arity 2, and repeats <expensive> for every call of the
function; the latter has arity 0, and shares <expensive>.  We don't want
to change behaviour.  Hence the call to exprIsCheap in ok_fun.

I noticed this when examining #18993 and, although it is delicate,
eta-reducing to a PAP happens to fix the regression in #18993.

HOWEVER, if we transform
   \x. f y x   ==>   f y
that might mean that f isn't saturated any more, and does not inline.
This led to some other regressions.

TL;DR currently we do /not/ eta reduce if the result is a PAP.