Align dummy column and APC column order (#3451)

In direct write to APC trace witgen, it's required that the order of
original columns in original instruction execution order should match
the order of post-optimization APC columns. This wasn't the case because
previously we use an `AllChildren` visitor that chains
`AlgebraicExpressions` in all constraints and bus interactions, whose
`AlgebraicReference` don't necessarily follow the order of
`StructReflection` of orignal airs columns.

However, thanks to `StructReflection`, `poly_id` of original air columns
are dispensed in the same order as the struct, and `globalize_reference`
also keeps the same order before optimization. Therefore, the only work
needed is to sort the unique columns returned by the `AllChildren`
implementation of `SymbolicMachine` in the same order as the `poly_id`
of original air columns.

This turns out to be quite efficient as well via a `BTreeMap`, as
explained in the comment below.

Author: qwang98

autoprecompiles/src/powdr.rs

@@ -39,6 +39,7 @@ pub trait UniqueReferences<'a, T: 'a> {
 }
 
 impl<'a, T: 'a, E: AllChildren<AlgebraicExpression<T>>> UniqueReferences<'a, T> for E {
+    // Output unique column references sorted by ascending id of original instruction columns
     fn unique_references(&'a self) -> impl Iterator<Item = AlgebraicReference> {
         self.all_children()
             .filter_map(|e| {
@@ -48,7 +49,9 @@ impl<'a, T: 'a, E: AllChildren<AlgebraicExpression<T>>> UniqueReferences<'a, T>
                     None
                 }
             })
-            .unique()
+            .map(|r| (r.id, r))
+            .collect::<BTreeMap<_, _>>()
+            .into_values()
     }
 }