Improve handling of the global recorder instance

[Noratrieb]

This gets rid of the dangerous static mut, adds more comments about the code, relaxes the orderings and documents the unsoundness of the clear function.

It implements a small once_cell-like abstraction.

I also noticed that the whole #[cfg(atomic_cas)] seems pretty pointless since this crate doesn't support no_std anyways.

About clear_recorder: This function was unsound before as well, since the following interleaving was possible:

• Thread A reads the state as init
• Thread B clears
• Thread B CASes the state and starts writing to the recorder
• Thread A now reads the recorder - a data race!

[tobz]

Everything checked out performance-wise. 👍🏻

Thank you for your contribution!

[tobz]

Released in metrics@v0.19.0 (along with other workspace crate releases to move to the new minor.)