Use linked list in ReleasePool

[kngwyu]

Fix #271
I haven't benched yet and maybe LinkedList<[NonNull<_>; 256] or so is more efficient

[kngwyu]

The 1st commit is naive LinkedList<NonNull<PyObject>> implementation.
The 2nd & 3rd commit is LinkedList<[NonNull<PyObject>; 256]> implementation.
I'll note a benchmark result for both two.

[kngwyu]

Benchmark result:
Vec<NonNull<_>>(before)

running 1 test
test iter_dict ... bench:   1,943,691 ns/iter (+/- 53,443)

LinkedList<NonNull>(after: 1)

running 1 test
test iter_dict ... bench:   6,349,135 ns/iter (+/- 259,604)

LinkedList<[NonNull<_>; 256>(after: 2)

running 1 test
test iter_dict ... bench:   2,172,151 ns/iter (+/- 248,869)

LinkedList<[NonNull<_>; 512]>(Not pushed)

running 1 test
test iter_dict ... bench:   2,325,197 ns/iter (+/- 110,542)

(Arch Linux 4.18.16, rustc 1.32.0-nightly (25a42b2ce 2018-11-07), Intel Core i5-6500)

[konstin]

Filed #282 because this is lurking UB

[konstin]

Big thanks for tackling this!

I wonder if there's a performance overhead in using a vec or zero-initialization. This would allow us to avoid mem::uninitialized()

[kngwyu]

@konstin
Benchmark result:
[T; 256]

running 1 test
test iter_dict ... bench:   2,222,297 ns/iter (+/- 69,061)

Vec built by Vec::with_capacity(256) (https://github.com/kngwyu/pyo3/tree/list-vec)

running 1 test
test iter_dict ... bench:   2,591,632 ns/iter (+/- 166,590)

(Ubuntu 17.10 + AMD Ryzen 5 1600)

[konstin]

Thanks for measuring, we now have a good rationale for keeping mem::uninitialized().