Optimize to_binary() function and to_bytes() method

[skirpichev]

It seems, that using mpn_get_str() is more efficient than generic mpz_export().

Some benchmarks are here:
#404 (comment)

Not sure what else we can do for #404. In the python-gmp I've added also the __reduce__ dunded method. This seems slightly better than rely on copyreg to support pickling:

Benchmark	ref	patch	gmp
dumps(1<<7)	23.9 us	23.8 us: 1.01x faster	22.6 us: 1.06x faster
dumps(1<<38)	24.0 us	23.9 us: 1.01x faster	22.7 us: 1.06x faster
dumps(1<<300)	24.1 us	23.8 us: 1.01x faster	22.9 us: 1.05x faster
dumps(1<<3000)	26.8 us	25.2 us: 1.07x faster	23.8 us: 1.13x faster
Geometric mean	(ref)	1.02x faster	1.07x faster

Can we add pickling to the gmpy2 with even less overhead? I don't know.

But if we avoid pickle machinery, you can see noticeable performance boost for small numbers too:

Benchmark	to_binary-ref	to_binary-patch
dumps(1<<7)	323 ns	300 ns: 1.08x faster
dumps(1<<38)	352 ns	315 ns: 1.12x faster
dumps(1<<300)	603 ns	436 ns: 1.39x faster
dumps(1<<3000)	3.17 us	1.57 us: 2.02x faster
Geometric mean	(ref)	1.35x faster

New code seems faster than int.to_bytes() roughly from 500bit numbers on my system.

benchmark code

# pickle-bench.py
import pyperf
from pickle import dumps
from gmpy2 import mpz


runner = pyperf.Runner()
ints = "1<<7", "1<<38", "1<<300", "1<<3000"

for v in ints:
    n = mpz(eval(v))
    v = "dumps(" + v + ")"
    runner.bench_func(v, dumps, n)

# pickle-bench2.py
import pyperf
from gmpy2 import mpz, to_binary


runner = pyperf.Runner()
ints = "1<<7", "1<<38", "1<<300", "1<<3000"

for v in ints:
    n = mpz(eval(v))
    v = "dumps(" + v + ")"
    runner.bench_func(v, to_binary, n)

[skirpichev]

N.B. this patch has two helpers, added to src/gmpy2_macros.h. I'm not sure if it's a right place or they correctly named.