Interleaving elaboration's `expose_boundary_tensor` is insufficiently general

Suppose we create a map with type `A + B` in dataflow.
This should be exposed in "control" as a map using the 'ambient tensor' (which uses `+`).

This is what `expose_boundary_tensor` does.
But currently, it is insufficiently general: it uses a simple pattern match on hexprs to find rightmost `+` operations, which are popped.
This fails in examples like below:

```
[i . i] ({a b} +) [j . j]
```

... where the `+` operation is "hidden" by a spider.

To implement this properly, one must interpret the type map as an open hypergraph, then remove `+` operations on the right boundary, replacing them with their sources.

See [here](https://github.com/hellas-ai/catena-lang/pull/4/changes#r3217067463)

Start at the expose_boundary_tensor entry point and review the linked pull request discussion. Model the type map as an open hypergraph, including the example where the + operation is hidden by a spider. Done means right-boundary + operations are removed and replaced with their sources in cases such as the supplied example.

Written by the indexing model from the issue text.