fix `string-normalize-nf[k]c' when composed is much smaller then decomposed

For composed normalization, space is initially allocated based on the
decomposed length. The clean up step to avoid wasted space was wrong.

collects/tests/racket/uni-norm.rktl

@@ -80,4 +80,10 @@
 	      (test c5 string-normalize-nfkd c5)))
 	  test-strings)
 
+;; Some tests where the composed form is much smaller than decomposed:
+(let ()
+  (define s (list->string (for/list ([i 64]) #\é)))
+  (test s string-normalize-nfc s)
+  (test s string-normalize-nfc (string-normalize-nfd s)))
+
 (report-errs)

src/racket/src/string.c

@@ -3891,8 +3891,8 @@ static Scheme_Object *normalize_c(Scheme_Object *o)
 
   s2[j] = 0;
   if (len - j > 16) {
-    s2 = (mzchar *)scheme_malloc_atomic((j + 1) * sizeof(mzchar));
-    memcpy(s2, s, (j + 1) * sizeof(mzchar));
+    s = (mzchar *)scheme_malloc_atomic((j + 1) * sizeof(mzchar));
+    memcpy(s, s2, (j + 1) * sizeof(mzchar));
     s2 = s;
   }