Add support for instanced subgraphs in ktdf_arch.

[KFAFSP]

We're starting to run into the issue that the device declaration files are becoming quite large to fit all the copies of groups of a kind. Since these copies must all be identical in the IR (required by having the same kind), these ops carry little information. Still, they must all exist as unique objects in memory to allow resource allocation.

This PR makes a few backwards compatible changes to ktdf_arch, including turning its containers into graph regions. It then adds the two new ktdf_arch.neighbor(hood) operations, which can be used to define regular recurrences of subgraphs.

This allows the user to write an architecture specification that lists every unique topology only once, and then parameterize over its repetitions. There is an example that features a ring topology. The resulting graph must be instantiated before it is consumed by the scheduler. This is done via the ktdfarch-instantiate-neighborhoods pass.

EDIT: As a side note, neighborhoods are regular subgraphs. That means existing analyses such as ResourceKinds will observe the same structure regardless of whether the neighborhoods were instantiated or not. They just won't be traversed by endpoint lookups, making each value that would cross their boundary opaque.

[KFAFSP]

@bmahjour Ive added ktdf-archgen and tested bytcode, and it actually already works out of the box. Import declarations can already import bytecode, and the tool can already write and recover it. Parsing it is also much faster - I did not add skipping the verifier to the importer though, that could be for later.

[bmahjour]

Please add a more complicated test involving nesting. I get a crash when I try the following:

ktdf_arch.device @nested {
  %outer = neighborhood %self_o : (exec_unit)[2] {
    %inner = neighborhood %self_i : (exec_unit)[3] {
      %e = exec_unit @exec
      yield %e
    }
    %e0 = neighbor affine_map<(d0) -> (d0+1)> in %inner : (exec_unit)[3]
    yield %e0
  }
  %a = neighbor affine_map<() -> (0)> in %outer : (exec_unit)[2]
  %b = neighbor affine_map<() -> (1)> in %outer : (exec_unit)[2]

  datapath %a to %b : exec_unit, exec_unit
}

It has to do with the (d0) -> (d0+1) map to select from the %inner. It goes through if I map (d0) -> (0).

[KFAFSP]

Yes, I will. I must have messed up the out of bounds check there, thanks for catching it!

[KFAFSP]

I was unable to reproduce a crash and instead observed only the expected behavior. I have added a test regardless. Maybe you were running an old commit?