Add optimized version of memchr

Created by: matijaskala

This is what musl does - read 8 bytes at a time until a match is found

It would be good to have some benchmarking to confirm that this improves performance, and show how much it helps.

Created by: matijaskala

#include <string.h>
#include <stdio.h>
#include <time.h>

void* memchr1 (const void* s, int c, size_t n);
void* memchr2 (const void* s, int c, size_t n);

char s[1000000];

int main () {
    memset (s, 0, sizeof(s));
    s[999999] = 1;
    struct timespec t1, t2, t3, t4;
    clock_gettime (CLOCK_REALTIME, &t1);
    for (int i = 0; i < 100; i++)
        memchr1 (s, 1, sizeof(s));
    clock_gettime (CLOCK_REALTIME, &t2);
    clock_gettime (CLOCK_REALTIME, &t3);
    for (int i = 0; i < 100; i++)
        memchr2 (s, 1, sizeof(s));
    clock_gettime (CLOCK_REALTIME, &t4);
    printf ("%ld\n%ld\n", (t2.tv_sec-t1.tv_sec)*1000000000l + t2.tv_nsec - t1.tv_nsec, (t4.tv_sec-t3.tv_sec)*1000000000l + t4.tv_nsec - t3.tv_nsec);
}

# ./a.out
66656610
10906751

Created by: matijaskala

Offset  Speedup
1       -41.72%
2       -33.13%
3       -24.33%
4       -31.56%
5       -12.84%
6       -15.76%
7       -4.64%
8       55.62%
9       60.17%
10      64.52%
12      65.68%
14      63.10%
16      129.31%
18      149.42%
20      128.14%
22      138.40%
24      173.81%
26      164.67%
28      169.54%
30      152.82%
32      331.13%
34      222.47%
36      205.11%
38      186.38%
40      273.06%

Looks like the tests are failing. You can rebase on master to get the build job for aarch64 to pass.