[RTG][Elaboration] Add support for 'scf.if' and 'scf.for'

[maerhart]

Did a quick profile of the following input:

func.func @dummy(%arg0: index) -> () {return}

rtg.test @test : !rtg.dict<> {
  %0 = index.constant 0
  %1 = index.constant 1
  %2 = index.constant 10000000
  %3 = index.constant 2
  %4 = scf.for %i = %0 to %2 step %1 iter_args(%a = %0) -> (index) {
    %5 = index.add %a, %3
    scf.yield %5 : index
  }
  func.call @dummy(%4) : (index) -> ()
}

Currently takes ~5.4 sec on my machine, with most of the time spent internalizing the intermediate integers.
Will follow-up with a PR that refactors the ElaboratorValue to only internalize complex values such as sets and bags, but not primitive integers.

[darthscsi]

LGTM. I think maybe we can be a bit simpler for some of the structure rewriting, but I'm not going to push on that right now.

[darthscsi]

A number of passes have assumed that the visitor is cheap to construct and thus embed this kind of state locally and create new visitors as needed. I don't know if this is relevant here.

[darthscsi]

I'm wondering if there is some alternate, forward-rewriting kind of approach:
elab(for (ind = start, end) { block; }) -> elab(let ind = start in block) ; elab(for(ind = start + 1, end) { block }))
let is essentially the clone block with mapping api.

[maerhart]

You mean actually cloning the MLIR ops? I intentionally tried to write this in a way that does not require any IR cloning/materialization because that's quite expensive. But from a readability perspective that would be better, I agree.