The old code was caching per-type struct field info. Instead,
cache type-specific encoding funcs, tailored for that
particular type to avoid unnecessary reflection at runtime.
Once the machine is built once, future encodings of that type
just run the func.

benchmark               old ns/op    new ns/op    delta
BenchmarkCodeEncoder     48424939     36975320  -23.64%

benchmark                old MB/s     new MB/s  speedup
BenchmarkCodeEncoder        40.07        52.48    1.31x

Additionally, the numbers seem stable now at ~52 MB/s, whereas
the numbers for the old code were all over the place: 11 MB/s,
40 MB/s, 13 MB/s, 39 MB/s, etc.  In the benchmark above I compared
against the best I saw the old code do.

R=rsc, adg
CC=gobot, golang-dev, r
https://golang.org/cl/9129044