We create thousands of AudioBuffer objects per minute in many cases,
local testing on Linux desktop reveals we only end up reusing about
80% of what we allocate, so add a memory pool which ensures 99%
reuse for common playback scenarios.

The AudioPlayback clockless pipeline integration test shows an
improvement in runs/s of ~4-20% depending on system load. This is
a simple change, so the gains seem worth it.

BUG=none
TEST=new AudioBuffer unittests, existing ones pass.

Review-Url: https://codereview.chromium.org/2788483003
Cr-Commit-Position: refs/heads/master@{#461308}