fix: compile Python reference-type equality to identity comparison

[fabiomadge]

Part of #6491. Dafny gives class types identity equality and the verifier reasons about them accordingly, but the Python backend compiled ==/!= on reference types to Python's ==/!=, which dispatch to the overridable __eq__ hook. An extern class overriding __eq__ therefore made verified reference (in)equalities evaluate incorrectly at run time:

var a := Box.Make(1);
var b := Box.Make(1);
assert a != b;              // verified
print a != b;               // printed false with a structural __eq__ extern

Generated datatype __eq__ methods compared class-typed fields the same way, so D(a, 0) != D(b, 0) was also verified-true but runtime-false.

Fix. Emit is/is not for EqCommon/NeqCommon at reference types (after DatatypeWrapperEraser.SimplifyType, as elsewhere), and identity comparisons for reference-typed fields in generated datatype __eq__ methods. This mirrors the discrimination the Java and C# backends already make at the same decision point (JavaCodeGenerator emits == (Object) for ref types; CsharpCodeGenerator emits == (object)). Python was the only backend where even direct equality was affected.

Non-regression checked: x == null, trait-typed comparisons, array identity, and the equality-sensitive comp/dafny0 compile tests across all targets. The regression test uses an extern with a structural __eq__ override, asserting both direct and datatype-field inequality survive compilation.

Known boundary, documented on the new EqualityOperator helper: equality at an abstract type parameter (e.g. inside method Eq<T(==)>(x: T, y: T), or a generic datatype's T-typed field) still compiles to == — the static type is not a reference type there. This matches the other backends, whose universal-equality fallbacks (java.util.Objects.equals, object.Equals) have the same hole; closing it needs runtime type information (the TypeDescriptor discussion in #6491).

Scope note: this intentionally does not address the collection-level unsoundness of #6491 (set<Box> collapsing etc.), which exists on several backends and needs the runtime collections to key on identity for reference types — a larger, cross-backend design discussed in the issue.

By submitting this pull request, I confirm that my contribution is made under the terms of the MIT license.