*** There is an API change here: the introduction of the
LatinOffset int in the RangeTable struct. ***

* Avoid checking Latin range multiple times for non-Latin runes.
* Use linear search when it is faster than binary search.

go test -calibrate runs the calibration for where the linear/binary
crossover should be.

benchmark                       old MB/s     new MB/s  speedup
BenchmarkFields            36.27        41.43    1.14x
BenchmarkFieldsFunc        36.23        41.38    1.14x

The speedup here is evenly split between the linear scans
and the LatinOffset change. Both are about 1.07x.

R=r
CC=bradfitz, gobot, golang-dev
https://golang.org/cl/6526048