gh-129813: Use PyBytesWriter in _json:_match_number_unicode

[maurycy]

The change is very similar to #138956

Initially, I thought that there's a speed up, but I mixed the JSON files, and it's the reverse:

Benchmark	json-main	json-pybytes
json_loads_1000_ints	41.4 us	47.3 us: 1.14x slower
json_loads_1000_floats	77.1 us	81.9 us: 1.06x slower
Geometric mean	(ref)	1.10x slower

which makes perfect sense given the conversion:

https://github.com/python/cpython/pull/138957/files#diff-efe183ae0b85e5b8d9bbbc588452dd4de80b39fd5c5174ee499ba554217a39edR1110

I mixed up the JSONs; my bad.

I'm surprised with the speed up:

Benchmark

The script:

import pyperf
import json

runner = pyperf.Runner()

num_elements = 1000

int_data = json.dumps(list(range(num_elements)))
runner.bench_func(f'json_loads_{num_elements}_ints', json.loads, int_data)

float_data = json.dumps([float(i) for i in range(num_elements)])
runner.bench_func(f'json_loads_{num_elements}_floats', json.loads, float_data)

Where main is 4e00e25

The environment:

% ./python -c "import sysconfig; print(sysconfig.get_config_var('CONFIG_ARGS'))"
'--enable-lto' '--with-optimizations'

sudo ./python -m pyperf system tune ensured.

• Issue: [C API] PEP 782: Add PyBytesWriter API #129813

[maurycy]

I'm closing it. I don't think it's making things better.