This fixes Trac #5341 and #5342.  The question is about
what to do when floating out of the RHS of a Rec-bound
function, when there's a FloatCase involved.  For FloatLets
they can join the Rec block, but FloatCases can't.  But
we don't want to mess with the arity (that was the bug).
So in this (rather exotic case) we push the FloatCase
back inside any lambdas.

See Note [Floating out of Rec rhss]. It's a slightly ugly fix, but I
can't think of anything better, and I don't think it has any practical
impact.