Improve MPZ_from/to_int() speed on PyPy

[skirpichev]

On PyPy3.11:

Benchmark	ref	patch
mpz(1 << 7)	19.4 us	15.0 us: 1.29x faster
mpz(1 << 38)	20.3 us	15.0 us: 1.36x faster
mpz(1 << 300)	22.4 us	23.4 us: 1.04x slower
mpz(1 << 3000)	43.3 us	44.3 us: 1.02x slower
int(mpz(1 << 7))	8.66 us	986 ns: 8.78x faster
int(mpz(1 << 38))	8.89 us	988 ns: 9.00x faster
int(mpz(1 << 300))	12.5 us	12.6 us: 1.01x slower
Geometric mean	(ref)	1.84x faster

Benchmark hidden because not significant (1): int(mpz(1 << 3000))

On CPython 3.13:

Benchmark	ref	patch
mpz(1 << 7)	348 ns	305 ns: 1.14x faster
mpz(1 << 38)	352 ns	318 ns: 1.10x faster
mpz(1 << 300)	712 ns	702 ns: 1.01x faster
mpz(1 << 3000)	2.51 us	2.51 us: 1.00x faster
int(mpz(1 << 7))	191 ns	196 ns: 1.03x slower
int(mpz(1 << 38))	232 ns	244 ns: 1.05x slower
int(mpz(1 << 300))	598 ns	601 ns: 1.00x slower
int(mpz(1 << 3000))	2.31 us	2.31 us: 1.00x slower
Geometric mean	(ref)	1.02x faster

benchmark

# bench.py
import pyperf
from gmp import mpz

runner = pyperf.Runner()
values = ['1 << 7', '1 << 38', '1 << 300', '1 << 3000']

for v in values:
    s = 'mpz('+v+')'
    i = eval(v)
    runner.bench_func(s, mpz, i)

for v in values:
    s = 'int(mpz('+v+'))'
    m = mpz(eval(v))
    runner.bench_func(s, int, m)