Change PyRegistry::get, getOrThrow, and contains to accept
PyIdentifier instead of raw std::string. Update Registry stubs
to accept Identifier[_T] | str.

Without this, passing str to PyRegistry methods (get, contains, etc.)
would raise TypeError since pybind11 doesn't auto-convert.

- Declare endstone-stubgen (from the EndstoneMC/stubgen GitHub source,
  which handles std::function callable annotations and equality
  overloads cleanly) and ruff as inline PEP 723 dependencies so
  `uv run scripts/stubgen.py` resolves them automatically.
- Strip the bogus `from . import endstone` that Pybind11ImportFix
  injects alongside the absolute import.
- Retype `NAME = Identifier(...)` class constants to
  `NAME: Identifier[<owner>] = "<key>"`, tracking the enclosing class
  by indentation so the generic type parameter survives and the bare
  string remains visible as documentation.
- Rewrite `endstone.Identifier` references in submodules to bare
  `Identifier` plus a `from endstone import Identifier` so the
  constant blocks read cleanly.
- Deduplicate exports so Identifier/Registry don't appear twice in
  `__all__`.

- Class constants like Dimension.OVERWORLD, ActorType.ZOMBIE,
  Enchantment.PROTECTION are now typed as `Identifier[Owner]` rather
  than plain `str`. Static type checkers reject passing
  `Dimension.OVERWORLD` where `Identifier[ActorType]` is expected.
- Registry lookups and `Identifier<T>`-typed parameters accept either
  an Identifier or a str at runtime; stubs show this as
  `Identifier[T] | str` via the io_name split.
- Expose Identifier, Registry, and BlockType through the lazy_loader
  forwarding in the relevant package __init__.py files so the runtime
  matches the stubs.