The encoding/binary.{Read,Write} methods use reflection which is slow.
This is ok for prototyping but has a serious performance impact.  In
each case, we know exactly what we're trying to write out, so we'll
allocate our own buffer and do the explicit conversion ourselves.

As an example of the performance improvement this makes, here are the
results from an (internal) benchmark:

benchmark          old ns/op    new ns/op    delta
BenchmarkDecode       427190       318845  -25.36%

Add support for decoding opGet ('g')

A few more small fixes to the encoder and decoder.

Dumb encoder

Add support for enough opcodes to read Graphite's Rendering stream

This adds support for enough opcodes to unpickle a graphite rendering
message, described at
http://graphite.readthedocs.org/en/0.9.10/render_api.html#pickle.

I also made some internal changes to consistently report errors to the
main decoding loop and renamed 'append' to 'push' so as not to conflict
with Go's builtin append keyword.  The unit tests still pass, and there
are now two tests for sample graphite pickles too.