This patch makes GHC's floating more robust, by allowing it
to float unboxed expressions of at least some common types.

See Note [Floating MFEs of unlifted type] in SetLevels.

This was all provoked by Trac #12603

In working this through I also made a number of other corner-case
changes in SetLevels:

* Previously we inconsistently use exprIsBottom (which checks for
  bottom) instead of exprBotStrictness_maybe (which checks for
  bottoming functions).  As well as being inconsistent it was
  simply less good.

  See Note [Bottoming floats]

* I fixed a case where were were unprofitably floating an
  expression because we thought it escaped a value lambda
  (see Note [Escaping a value lambda]).  The relevant code is
       float_me = (dest_lvl `ltMajLvl` (le_ctxt_lvl env)
                  && not float_is_lam)   -- NEW

* I made lvlFloatRhs work properly in the case where abs_vars
  is non-empty.  It wasn't wrong before, but it did some stupid
  extra floating.