Unnecessary Core.Box in closure referring to itself

[Keno]

Consider:

julia> function foo()
           back(a::Int) = 1
           back(a::Float64) = back(Int(a))
           back
       end
foo (generic function with 1 method)

julia> @code_lowered foo()
CodeInfo(
1 ─      back@_2 = Core.Box()
│   %2 = %new(Main.:(var"#back#3"), back@_2)
│        Core.setfield!(back@_2, :contents, %2)
│   %4 = Core.isdefined(back@_2, :contents)
└──      goto #3 if not %4
2 ─      goto #4
3 ─      Core.NewvarNode(:(back@_5))
└──      back@_5
4 ┄ %9 = Core.getfield(back@_2, :contents)
└──      return %9
)

This pattern is fairly common in ChainRules.jl, causing a number of unnecessary boxes there. I guess this is consistent with our usual scoping rules in that functions don't generally introduce their own name as an identifier in their body, so this needs to be captured from the outer scope, but perhaps lowering could recognize this situation?

[oxinabox]

I think this pattern is not uncommon.
Though ChainRules definately does that more than most.
It is the pattern i often use when writing recursive functions to traverse trees.
Where one dispatch handles leaf-nodes and one branches.
(and then maybe closing over some Vector that we are pushing results into)

[JeffBezanson]

Closing in favor of #53295.