arrow

ARROW-708: [C++] Simplify metadata APIs to all use the Message class, perf analysis

This doesn't produce a meaningful perf improvement, but it does remove a fair amount of code which is nice.

Here is an interactive FlameGraph SVG: https://www.dropbox.com/s/kp8i5r3j7i0em02/ipc-perf-20170324.svg?dl=0

![screenshot from 2017-03-24 12 52 54](https://cloud.githubusercontent.com/assets/329591/24304760/f283960a-1090-11e7-9bc5-4cb26f7ca0ae.png)

So it appears that out of the few hundred nanoseconds spent constructing each Array object, the time is mostly spent in object constructors. One thing that shows up is the RecordBatch constructor which is spending a bunch of time copying the `vector<shared_ptr<Array>>` passed, so I added a move constructor.

Author: Wes McKinney <wes.mckinney@twosigma.com>

Closes #437 from wesm/record-batch-read-perf and squashes the following commits:

95fdbc7 [Wes McKinney] Add RecordBatch constructor with rvalue-reference for the columns
793b3be [Wes McKinney] Inline SliceBuffer
212f17f [Wes McKinney] Benchmark in nanoseconds
a295aae [Wes McKinney] Remove record batch / dictionary PIMPL interfaces, handle flatbuffer details internally