This has about an order of magnitude lower runtime and/or call-overhead as
compared to the more generic `text-icu` approach, e.g. according to criterion
with GHC 7.4.1 on Linux/x86_64:
* 12 times faster for empty input strings,
* 6 times faster for 16-byte strings, and
* 3 times faster for 1024-byte strings.
`decodeLatin1` is also faster compared to using `decodeUtf8` for plain ASCII:
* 2 times faster for 16-byte input strings,
* ~40% faster for 1024-byte strings.