Decode as much as possible of a Huffman symbol in a single table
lookup (much like the zlib implementation), filling more bits
(conservatively, so we don't consume past the end of the stream)
when the code prefix indicates more bits are needed. This
results in about a 50% performance gain in speed benchmarks.
The following set is benchcmp done on a retina MacBook Pro:

benchmark                            old MB/s     new MB/s  speedup
BenchmarkDecodeDigitsSpeed1e4           28.41        42.79    1.51x
BenchmarkDecodeDigitsSpeed1e5           30.18        47.62    1.58x
BenchmarkDecodeDigitsSpeed1e6           30.81        48.14    1.56x
BenchmarkDecodeDigitsDefault1e4         30.28        44.61    1.47x
BenchmarkDecodeDigitsDefault1e5         32.18        51.94    1.61x
BenchmarkDecodeDigitsDefault1e6         35.57        53.28    1.50x
BenchmarkDecodeDigitsCompress1e4        30.39        44.83    1.48x
BenchmarkDecodeDigitsCompress1e5        33.05        51.64    1.56x
BenchmarkDecodeDigitsCompress1e6        35.69        53.04    1.49x
BenchmarkDecodeTwainSpeed1e4            25.90        43.04    1.66x
BenchmarkDecodeTwainSpeed1e5            29.97        48.19    1.61x
BenchmarkDecodeTwainSpeed1e6            31.36        49.43    1.58x
BenchmarkDecodeTwainDefault1e4          28.79        45.02    1.56x
BenchmarkDecodeTwainDefault1e5          37.12        55.65    1.50x
BenchmarkDecodeTwainDefault1e6          39.28        58.16    1.48x
BenchmarkDecodeTwainCompress1e4         28.64        44.90    1.57x
BenchmarkDecodeTwainCompress1e5         37.40        55.98    1.50x
BenchmarkDecodeTwainCompress1e6         39.35        58.06    1.48x

R=rsc, dave, minux.ma, bradfitz, nigeltao
CC=golang-dev
https://golang.org/cl/6872063