Add SensMixed/lmerTest-backed Mixed Assessor Model variant (rigorous F-test)

[kgdunn]

Background

PR #430 adds a pure-Python Mixed Assessor Model (MAM) in
process_improve.sensory.mam:

• mixed_assessor_model returns each panelist's scaling coefficient beta per
  attribute and a product-effect F-test that uses the genuine disagreement term
  as the error (more powerful than the classical interaction-based test), plus
  the classical F for comparison.
• align_scores harmonizes all panelists onto a common scale.

The F-test there is computed on replicate-averaged cell means with a
closed-form scaling/disagreement split. That is sufficient for the current
needs, but it is not the full random-effects treatment.

Proposed follow-up

Add an optional, more rigorous MAM variant backed by R's SensMixed /
lmerTest (the reference implementation, Brockhoff, Schlich & Skovgaard 2015,
Food Quality and Preference 39:156-166) via an rpy2 / Rscript bridge:

• proper mixed-model (REML) estimation with panelist as a random effect and
  replicate/session terms,
• the SensMixed scaling-vs-disagreement F-tests and automated random-effect
  selection,
• validate the pure-Python beta_i and F-tests against SensMixed on a
  published MAM dataset.

Keep it optional (behind an extra), so the pure-Python path remains the default
and the package has no hard R dependency.

Acceptance

• An opt-in MAM path that returns SensMixed-equivalent F-statistics and
  random-effect selection.
• A cross-check test comparing the pure-Python beta_i to the R output within
  tolerance.