Use a custom error type for invalid lengths, replacing `fmt.Errorf` (#69)

* Add benchmarks for different kinds of invalid UUIDs

Also add a test case for too-short UUIDs to ensure behavior doesn’t
change.

* Use a custom error type for invalid lengths, replacing `fmt.Errorf`

This significantly improves the speed of failed parses due to wrong
lengths. Previously the `fmt.Errorf` call dominated, making this the
most expensive error and more expensive than successfully parsing:

    BenchmarkParse-4                 29226529        36.1 ns/op
    BenchmarkParseBadLength-4         6923106       174 ns/op
    BenchmarkParseLen32Truncated-4   26641954        38.1 ns/op
    BenchmarkParseLen36Corrupted-4   19405598        59.5 ns/op

When the formatting is not required and done on-demand, the failure per
se is much faster:

    BenchmarkParse-4                 29641700        36.3 ns/op
    BenchmarkParseBadLength-4        58602537        20.0 ns/op
    BenchmarkParseLen32Truncated-4   30664791        43.6 ns/op
    BenchmarkParseLen36Corrupted-4   18882410        61.9 ns/op