I think we can fix the performance issues by raising the level at which allocation/free goes through a function pointer.

Instead of a malloc-like interface void *malloc(size_t size), we should be returning partially initialized objects.
PyObject *obj_malloc(PyTypeObject *tp, size_t size, size_t presize) would allocate a chunk of memory size + presize, returning a PyObject * pointing to that memory + presize, with the ob_type field set to tp and the ob_refcount set to one.

This is low-enough level to be fully general, but with enough context to support tracemalloc.

I think we would need the following implementations, switchable at runtime:

• default
• tracemalloc
• custom (used when the underlying PEP 445 allocator is changed)
• custom-tracemalloc (used when the underlying PEP 445 allocator is changed)
• freethreading
• freethreading-tracemalloc

We don't need (or want) to switch between the free-threading and default allocators, but it keeps the rest of the code simpler if they have the same interface.