Three compile perf improvements with deep nesting

These were changes are all triggered by #24471.

1. Make GHC.Core.Opt.SetLevels.lvlMFE behave better when there are
   many free variables.  See Note [Large free-variable sets].

2. Make GHC.Core.Opt.Arity.floatIn a bit lazier in its Cost argument.
   This benefits the common case where the ArityType turns out to
   be nullary. See Note [Care with nested expressions]

3. Make GHC.CoreToStg.Prep.cpeArg behave for deeply-nested
   expressions.  See Note [Eta expansion of arguments in CorePrep]
   wrinkle (EA2).

Compile times go down by up to 4.5%, and much more in artificial
cases. (Geo mean of compiler/perf changes is -0.4%.)

Metric Decrease:
    CoOpt_Read
    T10421
    T12425