Phase 2: PMI core metric and component reordering

[neuromechanist]

Part of #133.

Clean-room Python reimplementation of Pairwise Mutual Information (PMI) between
AMICA components, from the algorithm description only (binned marginal/joint
entropy with a Miller-Madow-style bias correction), NOT from postAmicaUtility's
GPL-2.0-or-later MATLAB source (minfojp.m/get_mi.m/arrminf2.m). Cite Palmer,
Ozgur Balkan, Arnaud Delorme, and Makoto Miyakoshi in the module docstring.

• Pairwise MI matrix across a model's components (histogram/entropy scheme,
  matching minfojp.m's bias-corrected estimator conceptually).
• Greedy block-diagonal component reordering (equivalent to arrminf2.m) so
  correlated components cluster near the diagonal.
• Tests: real bundled sample EEG only (NO MOCK data).