fix(hir): bare-factory parent's field initializers run on subclass instances

[proggeramlug]

Summary

Fixes the bare-factory field-init drop surfaced by PR #822's #806 harness (section 1). class Sub extends makeFactory() {} had extends_name = None and only extends_expr set, so the codegen-time chain walks (apply_field_initializers_recursive, the keys-array generator) walked by extends_name only, saw None, and skipped the factory class's field initializers entirely — new Sub().kind read undefined instead of the parent's kind = \"bare\" literal.

Minimal repro

function makeBare() {
  return class { kind = "bare"; extra = 42; hello() { return "hi"; } };
}
class Sub extends makeBare() {}
new Sub().kind;   // undefined before; "bare" after
new Sub().extra;  // undefined before; 42 after
new Sub().hello() // "hi" both before and after (methods already dispatched via CLASS_REGISTRY)

Fix

New HIR post-pass infer_dynamic_extends_names in crates/perry-hir/src/lower.rs. Runs after every function and class is in module (so forward-references work despite function hoisting). For each class with extends_name = None and extends_expr = Call(FuncRef(N)):

1. Look up function N's body.
2. Match Return(Some(ClassRef(name))) or Return(Some(Sequence([…, ClassRef(name)]))) (the latter is what PR fix(hir): mixin factories must register grandparent edge at runtime #826's mixin-grandparent fix emits when the factory's inner class itself has a dynamic parent).
3. If matched AND the returned class's field initializers are closure-free (no LocalGet anywhere — bare literals only), set extends_name = Some(name).

The closure-free guard is load-bearing. apply_field_initializers_recursive inlines each chained class's init expressions directly into the subclass's constructor IR, so an init like _tag = tag (where tag is the factory's parameter) would re-resolve LocalGet(tag) in the subclass's scope and read garbage. Without the guard, the tagged-factory section of the harness regresses. The guard conservatively skips those — only literal-initialized parents propagate. Method inheritance is unaffected: dispatch goes through the runtime CLASS_REGISTRY which the RegisterClassParentDynamic side effect populates regardless of this static fixup.

Test plan

• Bare-factory repro — Sub instances now report the parent's kind: bare, extra: 42 matching Node.
• tests: harness for dynamic class extension / mixins #806 harness — sections 1, 2 (tagged), 3 (captured-factory generic) now PASS. (Section 4 chained mixin still needs PR fix(hir): mixin factories must register grandparent edge at runtime #826; sections 5+ are follow-up work — super-args, static method dispatch, double-call factory _tag binding.)
• Class smoke set (test_simple_class, test_get_prototype_of_instance, test_issue_711_function_prototype, test_gap_class_advanced, test_class_static_iife) — all PASS byte-for-byte vs Node.
• cargo test --release -p perry-hir -p perry-codegen — all green.

Refs #321.