gh-139156: Use PyBytesWriter in the UTF-7 encoder

[vstinner]

Replace PyBytes_FromStringAndSize() and _PyBytes_Resize() with the PyBytesWriter API.

• Issue: Use PyBytesWriter in Unicode codecs #139156

[vstinner]

Benchmark:

import pyperf
runner = pyperf.Runner()
sizes = (3, 100, 1000)
for size in sizes:
    runner.timeit(f'{size:,} ASCII chars',
        setup=f's="x"*{size}',
        stmt='s.encode("utf7")')
for size in sizes:
    runner.timeit(f'{size:,} UCS-1 chars',
        setup=f's=chr(0xe9) * {size}',
        stmt='s.encode("utf7")')
for size in sizes:
    runner.timeit(f'{size:,} UCS-2 chars',
        setup=f's=chr(0x20ac) * {size}',
        stmt='s.encode("utf7")')
for size in sizes:
    runner.timeit(f'{size:,} UCS-4 chars',
        setup=f's=chr(0x10ffff) * {size}',
        stmt='s.encode("utf7")')

Results with CPU isolation and python -m pyperf system tune:

Benchmark	ref	pep782
3 ASCII chars	487 ns	480 ns: 1.01x faster
100 ASCII chars	785 ns	831 ns: 1.06x slower
1,000 ASCII chars	2.46 us	2.93 us: 1.19x slower
3 UCS-1 chars	499 ns	487 ns: 1.02x faster
1,000 UCS-1 chars	5.58 us	5.72 us: 1.03x slower
3 UCS-2 chars	498 ns	487 ns: 1.02x faster
100 UCS-2 chars	1.14 us	1.18 us: 1.04x slower
1,000 UCS-2 chars	6.01 us	6.04 us: 1.01x slower
100 UCS-4 chars	1.63 us	1.65 us: 1.01x slower
1,000 UCS-4 chars	10.4 us	10.4 us: 1.00x slower
Geometric mean	(ref)	1.02x slower

Benchmark hidden because not significant (2): 100 UCS-1 chars, 3 UCS-4 chars

I'm not sure what's going with 1,000 UCS-1 chars. It looks like a hiccup in the benchmark, not a real regression.

The change uses the same memory allocation strategy, so it should have basically no impact on performance.

[vstinner]

1,000 UCS-4 chars: 6.96 us => 6.48 us: 1.07x faster

If I re-run the benchmark with this change, I get: 5.84 us: 1.19x faster. Hum, it seems like the benchmark is not reliable :-(

[vstinner]

Benchmark on this change: python -m pyperf timeit -s 's=chr(0x10ffff) * 1000' 's.encode("utf7")'

• Build 1: Mean +- std dev: 5.95 us +- 0.06 us
• Build 2: Mean +- std dev: 5.72 us +- 0.08 us
• Build 3: Mean +- std dev: 6.34 us +- 0.51 us

New try with CPU isolation and python -m pyperf system tune:

• Build 1: Mean +- std dev: 10.4 us +- 0.0 us
• Build 2: Mean +- std dev: 10.4 us +- 0.0 us
• Build 3: Mean +- std dev: 10.4 us +- 0.0 us

[vstinner]

I'm not sure what's going with 1,000 UCS-1 chars. It looks like a hiccup in the benchmark, not a real regression.

If I re-run the benchmark, it becomes faster: Mean +- std dev: [ref] 486 ns +- 3 ns -> [pep782] 484 ns +- 4 ns: 1.01x faster