The output length only needs to be half the input length.

This uses criterions benchmark groups to compare the relative
performance between the different implementations at various byte
lengths. The criterion reports now provide graphs that overlay the
different implementation so you can easily see the relative performance.

Range patterns are deprecated, switch to inclusive ranges instead.
Re-export the deprecated function `hex_to` in a separate use statement
that allows deprecation warnings.

This isn't an error condition. A decoded empty slice is just an empty
slice.

This also removes functions from the public api. hex_decode and
hex_decode_unchecked are the only things exposed by default.
hex_check_sse, hex_check_fallback, and hex_decode_fallback are only
visible when compiled with the 'bench' feature. This means that benches
now need to specify --features=bench when running.

$ cargo bench --features=bench

This makes no changes to the actual implementation and benchmarks
confirm that.

Rather than checking to ensure all bytes are within valid ranges, this
now checks to ensure no bytes are within invalid ranges. This is mostly
to avoid compiler warning when comparing an i32 to 0xffff because we now
compare the i32 to zero.

The performance is identical.

…e pass over the input data.

This speeds up checked decoding substantially, while not changing the
performance of unchecked decoding.

On my machine this increases throughput when decoding 2 bytes of data by
20%, and when decoding 4096 bytes of data by 116%.

Use a lookup table to determine the offset to add to each byte. This
improves performance substantially.

decode/faster_hex/4096 thrpt increased 15%
decode/faster_hex_unchecked/4096 thrpt increased 54%

…an sse.

compare the implementation against the hex crate.

… of the crate.

The benches added unnecessary code that isn't available via the public
api and is only benchmarked as a side-effect of the decode benchmark.