bytes specialization, optimised from iterator implementation

[davidhewitt]

Similar to #4423, this is a followup to #4417 which sacrifices a bit of performance to instead stay away from unsafe code, while still trying to optimise things.

Relevant benchmarks before:

byte_slice_into_pyobject_small
                        time:   [8.0587 ns 8.1316 ns 8.2012 ns]

byte_slice_into_pyobject_medium
                        time:   [56.836 ns 57.127 ns 57.386 ns]

byte_slice_into_pyobject_large
                        time:   [6.8279 µs 7.0452 µs 7.2630 µs]

And after:

byte_slice_into_pyobject_small
                        time:   [4.6786 ns 4.7819 ns 4.9469 ns]

byte_slice_into_pyobject_medium
                        time:   [50.219 ns 50.611 ns 50.983 ns]

byte_slice_into_pyobject_large
                        time:   [7.4864 µs 7.5923 µs 7.7109 µs]

[davidhewitt]

Superseded by #4442