These tables are constant, so it makes sense for all encoders to share
one copy---but it was initialized in a racy way, and there's no
appealing way to fix that without adding dependencies on locking
functions. So we simply give each codec instance its own copy, which
costs 8 * (240 + 240 + 120 + 120) = 5760 bytes apiece.
As noted in the TODO comment, the size of the tables could be reduced,
and they could be filled in at compile-time, but that would make the
encoder output slightly different, which would mess with our tests.
R=henrik.lundin@webrtc.org, solenberg@webrtc.org
Review URL: https://codereview.webrtc.org/1177993003.
Cr-Commit-Position: refs/heads/master@{#9442}
ac81163011