gamut-png: entropy and bigram filter heuristics, with pruned two-tier trials

[justin13888]

gamut-png has one of the five filter-selection heuristics oxipng ships, and it is the one
oxipng demoted out of its own default preset.

Current state (src/filter.rs:26, src/encoder.rs:24-32): FilterStrategy::{None, Fixed, MinSumAbs, BruteForce}, where BruteForce tries six whole-image strategies, each fully
DEFLATEd at the final level, sequentially, with no pruning.

What is missing

technique	note
Entropy	min Shannon entropy of the filtered bytes. ~40 scalar lines.
Bigrams	fewest distinct byte bigrams. ~40 scalar lines.
BigEnt	min entropy over the bigram histogram.
Two-tier trial	compress trials at a cheap level, re-encode only the winner at the real level. Level::Fast already exists; there is no such tier.
AtomicMin pruning	a shared best-size-so-far, read before each trial, so a trial already over budget bails. Needs the deterministic tie-break tuple to keep output reproducible.
Per-line trial deflate	oxipng's -f 9: actually deflate 2–16 lines with prior context.

oxipng's preset table is the evidence for the first two: MinSum is dropped from every preset
except -o 0 and -o 6, and its default -o 2 is None, Sub, Entropy, Bigrams. That
demotion, by a project that benchmarks obsessively, is the strongest signal available — there are
no published byte tables, so gamut must measure it on its own corpus.

Negative result worth keeping

An oxipng contributor (an ECT author) reports that a genetic algorithm over filter choices —
pngwolf's actual method — "did not prove helpful." Do not build one.

Verification

benches/encode.rs prints the size table and tests/size_contract.rs gates it, both landed in
#224. A new heuristic that does not beat MinSumAbs on gamut's corpus should be recorded as a
negative result in STATUS.md rather than merged.

Refs #224