Add support for optimal renormalization.

Add support for optimal renormalization.

The current code for rounding probabilities down to a fixed resolution
is a bit too crude when resolution is low; whether it's optimal to round
up or down depends on the other frequencies, and the code for stealing
slots from other symbols also doesn't take this into account (as the
comment rightfully points out). These effects only really show up when
getting down to lower resolution, e.g. prob_bits = 10, but there, they
can be quite pronounced.

Add a function (0-clause BSD licensed for optimal usefulness, ie. public
domain without potential difficulty about whether private persons can put
anything in the public domain) to calculate the optimal distribution of
encoding slots, basically through brute force plus some memoization.
For e.g. the basic example (main.cpp), the number of bytes for prob_bits = 10
goes down from 440895 to 437590 bytes. (At the default prob_bits = 14, the
difference is very low; from 435113 to 435093 bytes.)

For main_alias.cpp, which has prob_bits = 16, I've kept the existing code,
so that it's not lost to the mists of time in case someone wants something
that's nearly even cheaper in terms of startup cost.