Compositional baseline featurizers (scale, AAC, ACC, …) + baseline comparison in AAPred

[breimanntools]

Problem

CPP produces positional Part–Split–Scale features. To show that CPP's positional
features actually beat simpler encodings, users need baseline featurizers — fixed-length,
non-positional sequence descriptors — to compare against (the use case's
"scale 61/60/75 vs CPP 71/74/94" barplot). Today:

• Only the scale-average baseline exists (SequenceFeature.scale_composition, née
  scale_mean). Other standard compositional/statistical descriptors are missing.
• Getting a baseline number is a manual in-between step: df_seq → X_baseline → model → score. The baseline is not a first-class capability of the class that does
  prediction/evaluation (AAPred), so every comparison re-implements the loop.

Goal

1. A small, consistent family of sequence-level baseline featurizers on
   SequenceFeature — each df_seq → X with no CPP positional split — covering the
   common descriptors used as ML baselines for protein sequences.
2. Baseline comparison built into the prediction class: AAPred.eval(..., baseline=...) computes the chosen baseline X internally from df_seq and reports
   its score as an extra row of df_eval, next to the CPP-features score — so the whole
   "CPP vs baseline" comparison (and aap.plot_eval barplot) comes from one call, with
   no manual featurization.

Scope / requirements

Baseline featurizers (each: df_seq + list_parts in, X (n_seq, n_dim) out;
return_df option; honest signatures; fully vectorized — no per-sequence Python loop;
non-canonical/gap handling documented; ≥30 tests incl. a byte-identical reference-value
test on a fixture):

• scale-based — scale_composition (mean scale value per scale over the span).
  Already shipped (feat(sf): SequenceFeature.scale_composition — scale-based baseline featurizer (#307) #319); this issue only tracks it as the first family member.
• amino-acid composition (AAC) — fraction of each canonical amino acid (20-dim).
• auto-cross-covariance (ACC) — lag-based auto/cross-covariance of scale values
  (captures short-range sequence order the pure composition drops); n_lag parameter.
• (candidates, decide during design) dipeptide composition (DPC), CTD, pseudo-AAC (PAAC).

Prediction-class integration:

• AAPred.eval(df_seq=..., baseline=<kind|True|list>) builds the baseline X
  internally and appends one df_eval row per baseline kind, using the same model/CV as
  the CPP evaluation (apples-to-apples). baseline=True ⇒ scale_composition (the
  default baseline).
• No change to the non-baseline df_eval shape/values when baseline is unset
  (additive).
• aap.plot_eval renders CPP vs baseline(s) as the comparison barplot.

Consistency: one naming scheme for the family (e.g. <basis>_composition +
acc); shared backend helpers; example notebook per method; docstrings that state the
application (which baseline, when to reach for it) up front.

Non-goals

• Not a general feature-engineering framework — a small, curated baseline set only.
• No new heavy dependencies (pure numpy/pandas; reuse the existing scales).
• Positional/CPP features stay the job of CPP / feature_matrix.

Measurable acceptance (KPIs)

• A single AAPred.eval(df_seq=..., baseline=True) reproduces the use-case baseline row
  (scale-average vs CPP across the reported metrics) — value-identical to the current
  manual scale_composition → model → score path.
• Each new descriptor: ≥30 unit tests, vectorized (assert no per-sequence Python loop in
  the hot path), and byte-identical to a documented reference implementation on a fixed
  input.
• AAPred.eval with baseline unset is byte-identical to today's df_eval; with
  baseline=[k1,k2] adds exactly len([k1,k2]) rows and nothing else.
• Each featurizer + the AAPred.baseline path is demonstrated in an executed example
  notebook.

Context

Follow-up to the epic #305 "collapse notebook boilerplate" work and the AAPred
prediction layer (#332). Grew out of the observation that scale_mean (now
scale_composition, #319) exists only to feed a baseline model, so the baseline belongs
where prediction/evaluation lives rather than as a manual featurization step.