string: Optimize memcmp

Rewrite memcmp to improve performance. On small and medium inputs
performance is typically 25% better. Large inputs use a SIMD loop
processing 64 bytes per iteration, which is 50% faster than the
previous version.