---------------------------------------------------------------------------
CM questions

1. a parallel update idea. We can add a data offset field to each
context, and set it on access. Then the update thread, running in
background, would just skip updates on contexts that got used again
before getting updated.

2. parallel CM stuff. When trying to thread it, it somehow always
turns into something like BWT - sorting the contexts, then
processing with a single model. But then, its only well defined for
typical order-N models.
2.1. Is there a way to do something similar to unordered set of
masked contexts and mixers (with contexts too)?
2.2. To be specific, is there a way to avoid coding chunk sizes and
just derive them from the data, like BWT does?

3. mixer fsm. We already know that a state machine is faster,
better, and uses less memory than a pair of bit counts or a
linear counter. We also more or less know how to build efficient
state machines, so that's ok.
But what about mixing? In theory (DMC tries to do it practically
though) it may be possible to make a dynamic state machine for the
whole data, but such a method would use statistics from a single
context anyway, which has its limits.
Anyway, preserving the state of previous occurrence of current o1
context within a single global state is just a waste of resources,
so its still necessary to maintain multiple states and mix them
somehow.
And that "somehow" is the problem, because the dumb solution like
s = T1[s1][s2][sm]; // s1,s2=context states, sm=mixer state
sm= T2[s1][s2][sm]; // update
would already require 32M at this point, which isn't possible for a
single mixer, even if we'd forget about initialization of these
tables.
So, basically, we need a mixer that would work on a function of
{s1,s2}, so that
s = T1[T3[s1][s2]][sm]
sm= T2[T3[s1][s2]][sm]
would work with 64k tables (which is still too much though).
Can we make such a version of logistic mixer?..