Add MIR/PMI to paper; trim to ~1500 words

[neuromechanist]

Follow-up to #201.

Added

• MIR/PMI: one sentence in the Implementation section noting pamica computes mutual information reduction (MIR) and pairwise mutual information (PMI), the separation-quality metrics used to benchmark ICA algorithms [@delorme2012independent] (they were shipped but unmentioned; AMICA.mir()/.pmi(), pamica/metrics/, docs in docs/api/metrics.md).

Trimmed 1769 -> 1493 words

Via an independent manuscript:paper-review (Sonnet) focused on cuts. JOSS papers target 250-1750 words, so this is well within range now. Cuts removed cross-section redundancy (the ~1e-15 score-function figure, "byte-identical EEGLAB I/O", and the float32/four-to-five-sig-digit fact were each stated 3-4 times; now stated once in their natural home), docs-level method detail (config-file specifics, version pins, per-run counts), and a table-narrating benchmark paragraph (Table 2 already carries the numbers). All load-bearing caveats kept (single subject, not core-matched, MLX-on-laptop, the multi-model iteration-budget limitation).

Fixed two accuracy issues the review caught

• Dangling reference: "The reference Fortran implementation shares these limitations" had no antecedent (the preceding sentence listed strengths); rewritten.
• Undefined symbol: $k$ was used six times but never defined; now defined at first use ("frames per squared channel count").

Body 1493 words, no em-dashes; Draft PDF CI validates the build.