feat: SequenceFeature.scale_mean — scale-average baseline featurizer

[breimanntools]

Part of #305.

Problem

The standard "scale-based vs CPP" comparison baseline — averaging each scale
over a whole sequence (no positional split) — has no library API, so the
γ-secretase notebook (cell 27) hand-rolls it with a raw residue comprehension:

def scale_X(df):
    seqs = (df["jmd_n"] + df["tmd"] + df["jmd_c"]).to_list()
    return np.array([df_scales_red.loc[[a for a in s if a in df_scales_red.index]].mean(axis=0).values
                     for s in seqs])

This is a legitimate, commonly-needed baseline featurization, but every user
reinvents it (and gets the missing-residue filtering subtly wrong).

Goal

Add a first-class scale-average featurizer that turns sequences + scales into a
(n_seq, n_scales) matrix, with the name matching its output noun.

Requirements

• Add SequenceFeature.scale_mean(df_seq, df_scales, list_parts=None)
  (or NumericalFeature if a better fit) returning the per-sequence scale
  average matrix. list_parts=None → whole sequence.
• Handle non-canonical / missing-in-scale residues consistently with the
  rest of the package; document the rule.
• Optional return_df=True for a labeled DataFrame (matches the package's
  return_df convention).
• numpydoc + per-method Examples include.

KPIs / Acceptance criteria

• Output equals the notebook's manual comprehension on the canonical fixture
  (within float tolerance).
• ≥1 unit test per parameter (positive + negative); covers an empty/all-
  non-canonical sequence edge case.

Scope / non-goals

• A no-positional-split mean only; positional splits remain CPP's job.

Dependencies

• Part of epic epic: use-case convenience methods — collapse γ-secretase notebook boilerplate #305 · relates CPP strategy preset (compositional/positional) convenience #87 (CPP strategy preset).

Standards checklist

• frontend/backend · validation block · numpydoc · tests · no-print