Profile-guided optimization (PGO) as optional part of builds

[N-Wouda]

The pydantic project had a bit of success adding PGO to their builds, as shown in the relevant PRs:

• build with PGO on x86_64 ubuntu and windows pydantic/pydantic-core#678
• PGO build for MacOS M1 pydantic/pydantic-core#1063
• add build-pgo make target pydantic/pydantic-core#741

I'm not sure if we'd see similar performance improvements from adding this to our build chain, but a 10-20% improvement might be doable and would not add much to actual release builds.

Let's investigate.

[N-Wouda]

Maybe relevant for quick benchmark checks? https://github.com/apps/codspeed-hq