Introduce vm.state.type_mutex to serialize type mutation and
version-tag assignment, matching CPython's free-threading model.

- Add with_type_lock() helper and split modified/assign_version_tag
  into inner (lock-free) and public (lock-acquiring) variants
- Add version_for_specialization() and lookup_ref_and_version_interned()
  for atomic type-cache lookups under the lock
- Wrap all type-mutating paths (__bases__, __annotations__, __module__,
  __type_params__, __doc__, SetAttr) with the type lock
- Drop old values outside the lock to prevent deadlock from weakref
  callbacks that may re-enter specialization
- Fix _ctypes set_attr calls to route through proper SetAttr path
- Reinitialize type_mutex after fork