Phase 1: MIR core metric (mutual information reduction)

[neuromechanist]

Part of #133.

Port mir(unmixing, data) -> (mir_nats, variance) from bigdelys/pre_ICA_cleaning's
getMIR.m (Apache-2.0, permits direct port with attribution) using pyAMICA's
existing sphere + unmixing pipeline.

• Use Seyed Yahya Shirazi's own mir() adaptation in
  sccn/NEMAR-pipeline/eeg_nemar_dataqual.m as the reference (it adds the
  sphering step), but fix the scope bug: that version's signature is
  (data, linT) yet its body references eig(W) and /N, which only resolve
  today via MATLAB's nested-function shared workspace. The Python port must use
  explicit parameters (linT, N = data.shape[1]), not replicate the implicit
  dependency.
• Add an attribution docstring/comment pointing to
  github.com/bigdelys/pre_ICA_cleaning/getMIR.m per the Apache-2.0 NOTICE
  requirement.
• Tests: real bundled sample EEG only (NO MOCK data), checking internal
  consistency (e.g. against a literal re-derivation of the marginal-entropy
  terms), not synthetic data.