We now track which control mode is used per byte and summarize the stats
in trace output when using new encoding.

Instead of just working on synthetic data, codecbench can now read a set
of files, recompress them into v0/v1 and profile decompression
performance.

This gives a more accurate representation vs just running codecbench, as
it's using the full data set in all its complexity, and instead of
rerunning one file many times it runs a sequence of files so the branch
predictor can't adapt too much to the specific structure of one file.

codecbench works better as a decoding benchmark since it can work on
real data set, whereas demo decoding may vary based on .obj file due to
various channels being redundant.

This makes it easier to validate output of codecbench vs codectest.

We use snake_case vs pascalCase.