Byte models for the arithmetic plane codecs.

A byte probability model the range-coding driver can run. Stateful: adaptive models mutate via [`Self::update`]; static models no-op.

Inverse of [`encode_bytes`]; produces exactly `decoded_len` bytes.

Drive `model` over `data`, returning the range-coded payload.

Order-0 model that starts uniform (every byte count = 1) and bumps the observed symbol's count after coding it. Rescales when the total nears [`TOTAL_CAP`]. No serialized state — encoder and decoder stay in lockstep.

Order-0 model with a fixed, normalized 256-bin count table.

Order-1 model: 256 independent adaptive sub-models keyed on the previous byte. Context starts at 0. Stateless on the wire.

Static order-1 (lag-1) model. Counts are fit once on the plane and frozen; [`update`] only advances the context. Contexts never observed as a predecessor fall back to the `marginal` (order-0) table. Every row is Laplace-smoothed and quantized to [`ORDER1_QUANT_TOTAL`], so no symbol is zero-width and the range coder never divides by zero.

Fixed PMF denominator cap. Kept well under the range coder's 2^24 renorm floor so `range / total >= 1` always holds.