optimize Py<T>::drop for the case when attached

[davidhewitt]

I was looking again at #5350 and realised it could be worth optimizing the Drop implementation of Py<T> to bias towards the case when the thread is attached, as I think this is a common scenario.

As part of the optimizations I split the code into a fast and #[cold] path, and also removed #[track_caller] as the panic only happens if the reference pool is not enabled.

[codspeed-hq]

CodSpeed Performance Report

Merging #5454 will improve performances by ×3

_{Comparing davidhewitt:py-drop (06f3235) with main (cb9111b)}

Summary

⚡ 2 improvements
✅ 96 untouched

Benchmarks breakdown

	Benchmark	BASE	HEAD	Change
⚡	collect_generic_iterator	105.4 ms	95.8 ms	+10.03%
⚡	drop_many_objects	11.4 µs	3.8 µs	×3