core: proto import types an outer FieldReference as the whole enclosing row instead of the referenced field

[nielspardon]

On proto → POJO import, an outer FieldReference is typed as the whole enclosing row, not as the field it selects. ProtoExpressionConverter (core/src/main/java/io/substrait/expression/proto/ProtoExpressionConverter.java:101 and :106) passes the scope from outerScopeForStepsOut / outerScopeForRelReference straight in as the knownType of newRootStructOuterReference / newRootStructOuterReferenceByRelReference. Root references don't have this problem: they go through StructField.constructOnRoot, which types the reference as struct.fields().get(offset). FieldReference.type() is documented as "the type of the referenced field".

With an enclosing row (name VARCHAR, s ROW(v INT)) and a reference to s:

FieldReference ref = FieldReference.newRootStructOuterReference(1, R.struct(R.I32), 1);
FieldReference back = (FieldReference) converterWithRoot.from(expressionProtoConverter.toProto(ref));
back.getType();                // Struct{Str, Struct{I32}}, the whole row, not Struct{I32}
ref.equals(back);              // false
back.dereferenceStruct(0);     // segments [SF(0), SF(1)] (s.v, correct), type Str (name's type)

Consequences:

• A POJO built from SQL (typed with the field type) never equals its proto round trip, so assertProtoPlanRoundrip / assertFullRoundTrip cannot be used on any correlated SQL, and the existing outer-reference round-trip tests declare the whole row as the reference type to match the import.
• getType() on an imported outer reference is wrong for any consumer that reads it, and since fix(core): retain reference scope during dereference #1268 made dereference* keep the outer scope, dereferencing one yields a correctly scoped reference with the wrong type, or throws for list/map dereferences. No in-repo path dereferences an imported outer reference today, but anyone adding nested outer-path support to the reader (rejected since fix(core): retain reference scope during dereference #1268) will hit this first.

The current behavior predates #1034, which corrected which scope is used but kept typing the reference as that scope. Was the whole-row type intended? If not, the fix is to type the reference as scope.fields().get(field) with the same bounds check as constructOnRoot, and to update the round-trip tests that encode the whole-row type. That changes what getType() returns on import, so it likely wants a fix(core)!. Related: #1095 (unvalidated knownType on the root-reference factory).

Measured on main at ca8315b.