Plugable bytereader

[shawjef3]

This builds upon #228 and is for #229.

before

Benchmark                    Mode  Cnt     Score     Error  Units
BenchmarkGet.withCaching     avgt   10  1082.168 ±  57.571  ns/op
BenchmarkGet.withoutCaching  avgt   10  6272.534 ± 155.222  ns/op

after

Benchmark                    Mode  Cnt     Score    Error  Units
BenchmarkGet.withCaching     avgt   10  1037.877 ± 72.202  ns/op
BenchmarkGet.withoutCaching  avgt   10  4412.954 ± 75.050  ns/op

[oschwald]

I haven't gone through this PR line by line and I haven't tested it myself, but I am a bit confused about why it would be faster after the change. It appears that we are primarily adding an abstraction layer, ByteReader, around the ByteBuffer, adding some casts, and some addition bounds checks. Do you have a hypothesis about what would make it faster?