actually fuzz utf8proc_normalize_utf32 in the fuzz target

[shamsManna]

the fuzz target allocates its utf-32 scratch buffer as size >= 4 ? NULL : malloc(size), so copy is only non-null when the whole input is under 4 bytes, and the very next statement size /= 4 then drives the length to 0. every one of the six utf8proc_normalize_utf32 calls has run with length 0 since the target was added, so a public entry point that consumes untrusted utf-32 has effectively had no fuzz coverage. flip the guard to size >= 4 ? malloc(size) : NULL and copy size * sizeof(utf8proc_int32_t) bytes so normalize reads real, initialized code points. locally 0 of 6160 normalize calls had a non-zero length before this, all of them do after, and the corpus runs clean under asan/ubsan.

[stevengj]

Thanks!