Fix to_builtins with frozendict - #1105

LGTM, thanks for the quick follow-up! Did a deep pass on 5a54d59 including an empirical run on 3.15.0a7.

Verified:

• The #1104 repro is fixed: enum keys/values and nested frozendicts convert in both modes, str_keys and order are honored now, plain-dict parity holds for content and ordering in all three order modes, and the type is preserved only with builtin_types=(frozendict,) - matching the flag's intent from #1052.
• Refcounting is clean on every path: _PyFrozenDict_NewSteal's NULL-guard covers the to_builtins_dict error paths (enc_hook raise, unsortable keys with order, str_keys TypeError), and 100k-300k iteration loops over both happy and error paths show zero RSS/gc-object growth.
• No dict-path regression (byte-identical behavior for plain dict, with and without the flag), encoders/convert untouched as they should be, builds and full suite green on 3.12 (PY315_PLUS off) and 3.15, including the PR diff re-applied onto current main.

Three small notes, none blocking:

1. In test_frozendict_builtin_types_sorted the assert_eq helper only checks ordering under if isinstance(left, typ) - if the fix ever regressed to returning a plain dict, equality would still pass (frozendict == dict) and the order check would silently skip, so the test would stay green without testing anything. An assert isinstance(left, typ) (or a type(res) is frozendict pin in the test) closes that - same lesson as the test_order_frozendict fix in #1107.
2. The PR body has no Closes #1104, so the issue won't auto-close on merge - worth adding.
3. Optional: one test line for a frozendict subclass input (it now collapses to exact frozendict/dict, consistent with dict-subclass behavior, but nothing pins that).