Implement faster memcmp for x86_64

[Demindiro]

On x86_64 it is generally fast to read unaligned words, so there is no need to do byte-by-byte comparisons.

The new algorithm compares 32 bytes at a time. If any doesn't match it effectively does a binary search to find the first non-matching byte.

Benchmark results on a Ryzen 7 2700X:

master (with extra benchmarks)

test memcmp_rust_1048576              ... bench:     293,975 ns/iter (+/- 3,188) = 3566 MB/s
test memcmp_rust_16                   ... bench:           7 ns/iter (+/- 0) = 2285 MB/s
test memcmp_rust_32                   ... bench:          11 ns/iter (+/- 1) = 2909 MB/s
test memcmp_rust_4096                 ... bench:       1,049 ns/iter (+/- 203) = 3904 MB/s
test memcmp_rust_64                   ... bench:          32 ns/iter (+/- 4) = 2000 MB/s
test memcmp_rust_8                    ... bench:           4 ns/iter (+/- 1) = 2000 MB/s
test memcmp_rust_unaligned_1048575    ... bench:     276,654 ns/iter (+/- 78,924) = 3790 MB/s
test memcmp_rust_unaligned_15         ... bench:           8 ns/iter (+/- 0) = 2000 MB/s
test memcmp_rust_unaligned_31         ... bench:          11 ns/iter (+/- 2) = 2909 MB/s
test memcmp_rust_unaligned_4095       ... bench:       1,234 ns/iter (+/- 301) = 3319 MB/s
test memcmp_rust_unaligned_63         ... bench:          33 ns/iter (+/- 11) = 1939 MB/s
test memcmp_rust_unaligned_7          ... bench:           4 ns/iter (+/- 0) = 2000 MB/s

memcmp-x86_64

test memcmp_rust_1048576              ... bench:      24,682 ns/iter (+/- 240) = 42483 MB/s
test memcmp_rust_16                   ... bench:           3 ns/iter (+/- 0) = 5333 MB/s
test memcmp_rust_32                   ... bench:           4 ns/iter (+/- 0) = 8000 MB/s
test memcmp_rust_4096                 ... bench:         113 ns/iter (+/- 1) = 36247 MB/s
test memcmp_rust_64                   ... bench:           5 ns/iter (+/- 0) = 12800 MB/s
test memcmp_rust_8                    ... bench:           3 ns/iter (+/- 0) = 2666 MB/s
test memcmp_rust_unaligned_1048575    ... bench:      27,049 ns/iter (+/- 6,989) = 38765 MB/s
test memcmp_rust_unaligned_15         ... bench:           3 ns/iter (+/- 0) = 5333 MB/s
test memcmp_rust_unaligned_31         ... bench:           4 ns/iter (+/- 0) = 8000 MB/s
test memcmp_rust_unaligned_4095       ... bench:         100 ns/iter (+/- 6) = 40960 MB/s
test memcmp_rust_unaligned_63         ... bench:           5 ns/iter (+/- 0) = 12800 MB/s
test memcmp_rust_unaligned_7          ... bench:           3 ns/iter (+/- 0) = 2666 MB/s

builtin

test memcmp_builtin_1048576           ... bench:      23,202 ns/iter (+/- 1,283) = 45193 MB/s
test memcmp_builtin_16                ... bench:           4 ns/iter (+/- 0) = 4000 MB/s
test memcmp_builtin_32                ... bench:           3 ns/iter (+/- 1) = 10666 MB/s
test memcmp_builtin_4096              ... bench:          73 ns/iter (+/- 16) = 56109 MB/s
test memcmp_builtin_64                ... bench:           4 ns/iter (+/- 1) = 16000 MB/s
test memcmp_builtin_8                 ... bench:           4 ns/iter (+/- 1) = 2000 MB/s
test memcmp_builtin_unaligned_1048575 ... bench:      25,353 ns/iter (+/- 793) = 41359 MB/s
test memcmp_builtin_unaligned_15      ... bench:           5 ns/iter (+/- 0) = 3200 MB/s
test memcmp_builtin_unaligned_31      ... bench:           4 ns/iter (+/- 0) = 8000 MB/s
test memcmp_builtin_unaligned_4095    ... bench:          85 ns/iter (+/- 4) = 48188 MB/s
test memcmp_builtin_unaligned_63      ... bench:           4 ns/iter (+/- 0) = 16000 MB/s
test memcmp_builtin_unaligned_7       ... bench:           4 ns/iter (+/- 0) = 2000 MB/s

Note that the results may vary a lot, e.g. memcmp_rust_unaligned_1048575 can be as slow as 26933 MB/s at times.

I also added some extra tests though I don't think there are enough yet. I am not sure what edge cases should be considered though.