err: don't use with_gil internally in PyErr::new()

[davidhewitt]

This is a partial follow-on from #1683

If we're going to remove the optimization of using PyO3's thread-local to check for the GIL being held in with_gil, then it seems desirable that PyErr::new really doesn't need the GIL at all.

I added a benchmark to accompany the change; it only slightly impacts performance (<2% slower and this is on the error path anyway). It also has the nice side effect that users already creating PyErr outside of a GIL will get a significant speedup, because the GIL won't be temporarily acquired.

[messense]

Warning: Performance alert! Previous value was 5255746.530213239 and current value is 2283663.643596061. It is 2.301453869947796x worse than previous exceeding a ratio threshold 2
Warning: Performance alert! Previous value was 6427231.592283239 and current value is 2608965.9845569944. It is 2.4635168224987765x worse than previous exceeding a ratio threshold 2

This action is dumb. 😅 Looks like it's intended for pytest for whatever reasons: https://github.com/rhysd/github-action-benchmark/blob/a1914d7dcbe14d006e4b5f12c7ff303a82a411f1/src/write.ts#L68

[davidhewitt]

Heh indeed, it could use some work on output formatting! I would write a PR myself but it looks like the maintainer isn't looking after it, and I don't want to devote time to maintaining a fork... https://github.com/rhysd/github-action-benchmark/pulls

[davidhewitt]

Anyone got any comments or concerns here, or should I look to merge this?

[davidhewitt]

Proceeding to merge this!