Remove unsound return of borrowed objects

[davidhewitt]

As discussed in #883, I think these are a few cases where we should be converting borrowed objects to owned objects for safety reasons.

Closes #883.

[davidhewitt]

In case we're worried about performance regressions, I made some benchmarks (from #891). I ran them on master, this PR, #887 , and also #887 with this PR cherry-picked on top.

Results:

	master	no-borrowed-objects	new-nativetypes	new-nativetypes + no-borrowed-objects
dict_get_item	3160 +- 820	3120 +- 750	2280 +- 1600	2530 +- 1260
iter_dict	3680 +- 1000	3720 +- 700	1620 +- 700	2330 +- 560
iter_list	2090 +- 550	2340 +- 1300	1080 +- 590	1610 +- 430
list_get_item	1080 +- 410	1100 +- 400	470 +- 230	740 +- 320
drop_many_objects	4.4 +- 1.2	4.8 +- 0.4	3.8 +- 0.4	3.7 +- 0.9
iter_tuple	780 +- 110	760 +- 560	790 +- 300	820 +- 340
tuple_get_item	960 +- 250	950 +- 100	420 +- 200	580 +- 330

Conclusions:

• This PR does now slow down vs master much (maybe just a little)
• new-nativetypes is much much faster than master (yay!)
• Applying this PR after new-nativetypes causes greater slowdown because the difference between borrowed and owned objects is greater after that. But the combination is still faster than master 😄

So I would be really keen to see this safety gap closed. If we are really worried about performance we could add unsafe fn get_item_borrowed to PyList etc., but I don't think the performance difference is so great that it's worth it.

[kngwyu]

Good catch, thank you!

[davidhewitt]

I've undone the change for PyTuple::get_item

[kngwyu]

Thank you!