perf: avoid repeated pointer traversals

Supposedly gives us a 0.5% win in CPU time:

    Summary of cpu time and (wall time):

                        best    avg      sd     +/-      p     (best)    (avg)      (sd)     +/-   p
         pathological 0.20469 0.21396 0.06910 [-0.5%]        (0.20469) (0.21400) (0.06916) [-0.5%]
            command-t 0.15212 0.15898 0.05456 [+1.4%]        (0.15211) (0.15898) (0.05456) [+1.4%]
    chromium (subset) 1.12325 1.18233 0.08018 [-0.9%]  0.025 (0.25472) (0.27047) (0.02438) [-0.1%]
     chromium (whole) 0.90673 0.91697 0.05037 [-0.2%]  0.005 (0.10008) (0.10497) (0.01930) [+0.2%]
           big (400k) 1.38030 1.38569 0.01572 [-0.6%] 0.0005 (0.14990) (0.15333) (0.00925) [-0.8%]
                total 3.77224 3.85793 0.13443 [-0.5%]  0.005 (0.86813) (0.90174) (0.10183) [-0.0%]

Author: wincent

lua/wincent/commandt/lib/score.c

@@ -35,11 +35,11 @@ static float recursive_match(
 ) {
     float *memoized = NULL;
     float seen_score = 0.0f;
+    const char *haystack_contents = m->haystack->candidate->contents;
 
     // Iterate over needle.
     for (size_t i = needle_idx; i < m->needle_length; i++) {
         char needle_char = m->needle_p[i];
-        const char *haystack_contents = m->haystack->candidate->contents;
 
         // Iterate over (valid range of) haystack.
         for (size_t j = haystack_idx; j <= m->rightmost_match_p[i]; j++) {