Migrate to pyo3 v0.21 Bound API

[MarshalX]

they allowed to manage memory from binding side instead of pyo3 internals (gif refs pool). i expect perf boost here

[codspeed-hq]

CodSpeed Performance Report

Merging #23 will improve performances by 54.73%

_{Comparing pyo3-0-21 (fa1aadf) with main (8365eab)}

Summary

⚡ 6 improvements
✅ 186 untouched benchmarks

Benchmarks breakdown

	Benchmark	main	pyo3-0-21	Change
⚡	test_dag_cbor_decode_fixtures[array-6433713753386423]	30.4 µs	26.6 µs	+14.19%
⚡	test_dag_cbor_decode_fixtures[dagpb_simple_forms_3]	101.1 µs	87.8 µs	+15.13%
⚡	test_dag_cbor_encode_real_data[canada.json]	73.3 ms	56.2 ms	+30.46%
⚡	test_dag_cbor_encode_real_data[citm_catalog.json]	118.2 ms	76.4 ms	+54.73%
⚡	test_dag_cbor_encode_real_data[github.json]	4.2 ms	3.8 ms	+11.01%
⚡	test_dag_cbor_encode_real_data[twitter.json]	36.9 ms	32 ms	+15.34%

[MarshalX]

perf boost in decoding:

degradation in encoding. probably i forgot to migrate something:

[LilyFoote]

Looks to me like you missed &PyAny in:

python-libipld/src/lib.rs

Line 160 in 901a716

fn encode_dag_cbor_from_pyobject<'py, W: Write>(py: Python<'py>, obj: &'py PyAny, w: &mut W) -> Result<()> {

and

python-libipld/src/lib.rs

Line 365 in 901a716

fn encode_multibase(code: char, data: &PyAny) -> PyResult<String> {

[MarshalX]

@LilyFoote thank you! now it much much better! still have 1 regress but ig because of new allocs in sort_map_keys. previously it used refs only. overall perf increased!

decode increased in all cases 🎉🎉:

encode:

[MarshalX]

I found this in PyO3/pyo3#3684

Update py.None(), py.Ellipsis() and py.NotImplemented() to return Borrowed<'py, 'py, PyNone> (etc.)

So should I update py.None() somehow?

[davidhewitt]

No need; early in 0.21 development I changed those APIs to return GIL Refs e.g. &PyNone. But when the Bound API became the main focus of 0.21 I reverted that change, they just return PyObject again which is fine for now.

[MarshalX]

new record! +52.9%. ive returned back to main branch and added profile-guided optimization:

• Test PGO on x86_64 ubuntu #21
• Add PGO-optimized wheels #30