[Arc] Add LowerCoroutines pass

[fabianschuiki]

Add an arc-lower-coroutines pass that converts coroutine definitions into plain state machine functions. The lowered functions take an explicit state and program counter as leading arguments, dispatch on the PC to the entry block or one of the resume blocks, and persist all values live across suspension points in a union with one struct variant per resume block.

Each definition is lowered independently: the values live across suspension points are first captured as trailing resume block arguments through a pruned SSA construction based on liveness and iterated dominance frontiers, with constants cloned into their using blocks instead of being persisted. A global sweep then rewrites coroutine calls into regular function calls, lowers the sentinel value ops to constants and comparisons, and replaces all occurrences of the opaque coroutine state and PC types -- including in unrelated ops and nested aggregates -- with the concrete union and integer types. Recursive coroutines are rejected, since their state would require unbounded storage.

Assisted-by: Claude Opus 4.8

[circt-bot]

Results of circt-tests run for 5aa4099 compared to results for 6a3a01e: no change to test results.

[fzi-hielscher]

I won't be able to do a full review today, just one first thought: Do we want to pass the state to the callee by-value?
I could imagine that the state union may get quite large in practice, even though, for a given suspension point, substantial portions can be unused. So, passing it by-value could lead to a lot of unnecessary copying and pressure on the registers.
On the other hand, I suspect the state SSA value will rarely remain live beyond a call to it's coroutine. So, if we pass it by reference instead, we could reuse the buffer for the new returned state.