This tunes the entropy model for coarse energy introduced in commit
c1c40a76.
It uses a constant set of parameters, tuned from about an hour and a
half of randomly selected test data encoded for each frame size,
prediction type (inter/intra), and band number.
These will be slightly sub-optimal for different frame sizes, but
should be better than what we were using.
For inter, this saves an average of 2.8, 5.2, 7.1, and 6.7 bits/frame
for frame sizes of 120, 240, 480, and 960, respectively.
For intra, this saves an average of 1.5, 3.0, 4.5, and 5.3 bits/frame
(for the same frame sizes, respectively).
This changes how the PDF used to code coarse energy.
New features:
1) The probability of 0 (p0) is now indepedent of the decay rate
of the remaining values; this additional flexibility will allow
us to model the actual distribution better, though that
improvement is not part of this patch.
2) There is a guaranteed minimum number of encodable energy
deltas.
This ensures that even the most extreme sudden volume changes
can be accurately represented.
3) The tail end of the distribution has an adjustable (through a
constant in the code) minimum probability.
This allows us to lower the worst-case bit cost of a single
delta.
4) The codebook is interleaved as 0, -1, +1, -2, +2, ... instead
of the 0, +1, -1, +2, -2, ... order used before (see 5).
5) There is no restriction that p0 be even.
Any remaining, unused part of the code is assigned to an
additional negative value (collected inter data suggests that
very large negative deltas are more common than very large
positive ones).
If the minimum probability is greater than 1, then an
additional positive delta with a smaller probablity may also
be added.
6) Once the tail of the distribution is reached, the energy delta
is computed directly, instead of continuing to loop through
the codebook.
This reduces the worst-case computational cost.
