always implement tp_traverse to traverse type object

[davidhewitt]

This is a followup to #6206 which I noticed while reviewing (also very similar to #6224).

Reading the CPython docs for tp_traverse:

Heap types (Py_TPFLAGS_HEAPTYPE) must visit their type with:

Py_VISIT(Py_TYPE(self));

All #[pyclass] types are heap types, but we were not doing that. The test added demonstrates a memory leak present on main which this branch fixes.

Summary of changes:

• We now make every PyO3 type a GC type, and install a tp_traverse which traverses the heap type
• Because every type is now a GC type, I could simplify create_type_object to
  • We no longer need separate tp_dealloc / tp_dealloc_with_gc implementations
  • I moved the assertion that __clear__ requires a __traverse__ implementation to compile-time in macro code (with a UI test)
• I tidied up the PyVisit type implementation slightly at the same time

[Icxolu]

If we apply a #[track_caller] to this, it might clean up the error backtrace a bit.

[davidhewitt]

Good idea, pushed that and will merge, thanks!