sum() deallocates float subclass instances as exact floats in the complex fast path

Bug report

Bug description:

Bug report

Description

In builtin_sum_impl() (Python/bltinmodule.c), the complex-number fast
path tests items with PyFloat_Check() but releases them with
_Py_DECREF_SPECIALIZED(item, _PyFloat_ExactDealloc).

PyFloat_Check() also matches subclasses of float, whereas
_PyFloat_ExactDealloc is only valid for exact float objects. When a
float subclass instance reaches refcount 0 through this path, it is
deallocated as an exact float, bypassing its real tp_dealloc
(instance __dict__, weakref list, and GC bookkeeping). This can leak
memory and leave the cyclic GC in an inconsistent state.

The adjacent float fast path in the same function already guards the
identical specialized dealloc with PyFloat_CheckExact(); only the
complex branch uses the broader PyFloat_Check().

Reproducer

class F(float):
    pass

# complex start value selects the complex fast path
sum([F(1.0), F(2.0)], 0j)

Proposed fix

In the complex branch, use PyFloat_CheckExact(item) instead of
PyFloat_Check(item), so subclass instances fall through to the generic
PyNumber_Add path with a normal Py_DECREF. Numeric behavior is
unchanged.

Environment

• CPython: main branch (please confirm the line is still present)

CPython versions tested on:

CPython main branch

Operating systems tested on:

Linux

Linked PRs

• gh-150869
• gh-151063

Inizia in Python/bltinmodule.c, in builtin_sum_impl(), e confronta il percorso rapido per complex con il percorso rapido adiacente per float. Verifica il controllo specializzato della deallocazione per le sottoclassi di float e conferma che il reproducer venga completato senza che la sottoclasse venga deallocata erroneamente; le PR collegate gh-150869 e gh-151063 indicano che il lavoro potrebbe essere già in corso.