Improve .extend() performance

[bluss]

cc #101

I'm experimenting with different formulations. Both the existing and this pr's implementations can compile into a memcpy, like they should, if the input is a cloned slice iterator. The main difficulty seems to be to make a good benchmark. Without "black_box"es, the benchmarks compile out (and that's normally a good sign in itself, the code is then transparent to the optimizer) and with too many black box calls, the optimizations are disabled.

[bluss]

Not entirely happy with these benchmarks either, see code in the PR, but they seem fair(? please review)

 name 63 ns/iter 62 ns/iter diff ns/iter diff % extend_with_constant 294 (1741 MB/s) 1 (512000 MB/s) -293 -99.66% extend_with_range 426 (1201 MB/s) 289 (1771 MB/s) -137 -32.16% extend_with_slice 424 (1207 MB/s) 13 (39384 MB/s) -411 -96.93% extend_with_write 13 (39384 MB/s) 13 (39384 MB/s) 0 0.00%

obviously when extend_with_constant optimizes out it doesn't tell us so much, except that the new extend code is somehow more transparent to the compiler than the old.

[bluss]

Comparison with try_extend_from_slice shows that they both compile to memcpy:

test extend_from_slice ... bench: 14 ns/iter (+/- 1) = 36571 MB/s
test extend_with_slice ... bench: 13 ns/iter (+/- 1) = 39384 MB/s

extend_with_slice is the regular extend() used with a slice iterator (benchmark is in the PR).