Try to use PyLong_As/FromNativeBytes API for mpz <-> int conversion

[skirpichev]

Closes #313

[skirpichev]

At this point it looks worse:

$ pyperf compare_to export-ref.json export-patch.json
1<<7: Mean +- std dev: [export-ref] 163 ns +- 1 ns -> [export-patch] 164 ns +- 1 ns: 1.01x slower
1<<38: Mean +- std dev: [export-ref] 214 ns +- 2 ns -> [export-patch] 216 ns +- 2 ns: 1.01x slower
1<<300: Mean +- std dev: [export-ref] 633 ns +- 7 ns -> [export-patch] 804 ns +- 8 ns: 1.27x slower
1<<3000: Mean +- std dev: [export-ref] 2.32 us +- 0.01 us -> [export-patch] 3.17 us +- 0.02 us: 1.37x slower
1<<10000: Mean +- std dev: [export-ref] 6.92 us +- 0.18 us -> [export-patch] 9.66 us +- 0.02 us: 1.40x slower

Geometric mean: 1.20x slower
$ pyperf compare_to import-ref.json import-patch.json 
1<<7: Mean +- std dev: [import-ref] 256 ns +- 4 ns -> [import-patch] 251 ns +- 2 ns: 1.02x faster
1<<38: Mean +- std dev: [import-ref] 268 ns +- 10 ns -> [import-patch] 325 ns +- 5 ns: 1.21x slower
1<<300: Mean +- std dev: [import-ref] 629 ns +- 11 ns -> [import-patch] 1.29 us +- 0.02 us: 2.05x slower
1<<3000: Mean +- std dev: [import-ref] 3.16 us +- 0.02 us -> [import-patch] 3.69 us +- 0.04 us: 1.17x slower
1<<10000: Mean +- std dev: [import-ref] 9.56 us +- 0.39 us -> [import-patch] 9.90 us +- 0.05 us: 1.04x slower

Geometric mean: 1.24x slower

Details

# bench-export.py

import pyperf

from gmp import mpz

cases = ['1<<7', '1<<38', '1<<300', '1<<3000', '1<<10000']
runner = pyperf.Runner()
for c in cases:
    i = eval(c)
    m = mpz(i)
    runner.bench_func(c, int, m)

# bench-import.py

import pyperf

from gmp import mpz

cases = ['1<<7', '1<<38', '1<<300', '1<<3000', '1<<10000']
runner = pyperf.Runner()
for c in cases:
    i = eval(c)
    runner.bench_func(c, mpz, i)