perf: optimize HPACK decoding with 8-bit tables

[seanmonstar]

Previously we used a table of nibbles to each state of decoding. Instead, we now use a few byte-wide tables. Common codes fit in the first table, rarer codes are completed looking in the next tables.

On my low powered machine, these are the bench results:

Before:

test hpack::huffman::bench::decode_all_octets   ... bench:       5,476.74 ns/iter (+/- 567.07) = 106 MB/s
test hpack::huffman::bench::decode_header_value ... bench:         765.42 ns/iter (+/- 145.32) = 62 MB/s
test hpack::huffman::bench::decode_long_ascii   ... bench:         905.51 ns/iter (+/- 161.67) = 66 MB/s
test hpack::huffman::bench::decode_short_ascii  ... bench:         172.34 ns/iter (+/- 34.22) = 69 MB/s

After:

test hpack::huffman::bench::decode_all_octets   ... bench:       3,887.27 ns/iter (+/- 678.71) = 149 MB/s
test hpack::huffman::bench::decode_header_value ... bench:         661.78 ns/iter (+/- 88.98) = 72 MB/s
test hpack::huffman::bench::decode_long_ascii   ... bench:         724.07 ns/iter (+/- 93.60) = 82 MB/s
test hpack::huffman::bench::decode_short_ascii  ... bench:         152.48 ns/iter (+/- 13.02) = 78 MB/s

Which is about 13-41% improvement, depending.

cc #883