feat: Added typed method registration for py_class

[Kathryn-cat]

Add type method registration on py_class

feat(dataclasses): add @method decorator for FFI TypeMethod registration

Closes the gap where @py_class-decorated Python classes couldn't
expose user-defined methods to the FFI reflection table. Trait
references like $method:NAME in __ffi_ir_traits__ now resolve
cleanly against Python-defined methods, on parity with C++-defined
refl::ObjectDef<T>().def("name", ...).

Architecture

• tvm_ffi.method decorator stamps fn.__ffi_method__ = True on the
  decorated function (unwraps staticmethod to __func__ so both
  @method @staticmethod and @staticmethod @method work).
• _collect_py_methods in py_class.py widens beyond the existing
  _FFI_RECOGNIZED_METHODS allowlist: it now also picks up any
  callable in cls.__dict__ carrying the __ffi_method__ marker
  via _is_method_marked. The allowlist still routes TypeAttrColumn
  dunders (__ffi_repr__, __ffi_ir_traits__, etc.) through
  TVMFFITypeRegisterAttr unchanged.
• Marked methods register through the existing TypeMethod path
  (TVMFFITypeRegisterMethod) — same machinery the C++ side uses.
  The C++ printer's FindMethod (pyast_trait_print.cc:130-148) walks
  info->methods[] and discovers these entries identically to
  C++-defined methods, including ancestor-walking for inheritance.
• _validate_method_name rejects @method decoration on:
  • __ffi_* reserved prefix (would double-register as TypeAttrColumn),
  • Python protocol dunders (__len__, __iter__, etc. — reserved
    for Python semantics),
  • @classmethod (the cls first arg breaks the packed-call
    convention; raises at decoration time).
• Identity-typed: @method returns the original function unchanged
  (no wrapping); IDE / type-checker sees the method's signature
  unmodified.

Public Interfaces

• tvm_ffi.method (re-exported from tvm_ffi.dataclasses)

Behavioral Changes

None for existing classes. Only NEW @method-decorated callables land
in TVMFFITypeInfo.methods[].

Tests

• tests/python/test_typed_method.py (new, 13 tests): registration
  shape, FFI Function callability, validation errors.
• tests/python/test_ir_traits.py (5 new tests): end-to-end via
  pyast.to_python() against $method: refs in AssignTraits.text_printer_post,
  `

[gemini-code-assist]

Code Review

This pull request introduces the @method decorator for explicit FFI TypeMethod registration, enabling C++ and Rust to resolve Python methods by name. It includes validation for reserved names and method types. Feedback suggests improving the decorator's robustness for callables with restricted attribute access and adding a redundant check for classmethods during registration to ensure FFI compatibility.

[gemini-code-assist]

Setting attributes on arbitrary callables might fail for certain types (e.g., built-in functions or objects with __slots__ that don't include __dict__). While @method is primarily intended for user-defined functions in a class body, adding a check or a try-except block would make the decorator more robust against unexpected input types.

[gemini-code-assist]

The logic for determining is_static and func assumes that any callable not wrapped in staticmethod is an instance method. However, if a user manually marks a classmethod (bypassing the decorator's check), is_static will be False and func will be the classmethod object itself. This might lead to unexpected behavior in the FFI packed-call convention which expects either a bound method or a plain function. Consider explicitly checking for and rejecting classmethod objects here as well, or ensuring the FFI layer handles them correctly.

Suggested change

	is_static = isinstance(value, staticmethod)
	func = value.__func__ if is_static else value
	is_static = isinstance(value, staticmethod)
	if not is_static and isinstance(value, classmethod):
	raise TypeError(f"@py_class({cls.__name__}): classmethod {name!r} is not supported for FFI registration.")
	func = value.__func__ if is_static else value

[Kathryn-cat]

Addressed in latest commit

[junrushao]

Please fix the CI, and update docstring to be decoupled with not-yet-available features

[junrushao]

LGTM!