Devirtualization of CallIndirect -> Call

[croyzor]

Useful for both BRAT and guppy.

We already have InlineCall rewrite (all Calls have a static known target). Moreover, dataflow analysis is already capable of identifying when the called-function input to CallIndirect has a single/known target. (For example, when the input of the CallIndirect is a LoadFunction, but also other cases.) So, extend constant folding that, when it sees the latter, converts the CallIndirect to a Call.

---

See #978. For BRAT compilation, we'd like to be able to inline static functions, so an indirect_call of a static graph coming from a LoadFunction node becomes an instantiation of that graph. We can then get further by constant folding the whole hugr.

[acl-cqc]

I think that's two separate steps - firstly IndirectCall + LoadFunction -> direct Call (which doesn't require inlining, but removes the need for indirect-call which will likely move to the tierkreis extension); secondly, inlining of arbitrary direct Calls...

[doug-q]

This would likely help lazify-measurement. Indirect calls are harder to change arguments of. Quantinuum/tket2#432

[acl-cqc]

So #2059 has done the groundwork and the analysis here; we just need to either add a new pass, or extend the constant-folding pass (maybe with a config), to convert CallIndirect when the target is known into direct Calls. (interpret_leaf_op gets all the info it needs - this is not exposed to extension ops yet (#2087) but we don't need that here.)