Add head-resolved four-head nanoGPT spectral analysis

[charlesmartin14]

Summary

• adds a standalone post-hoc analyzer for the completed four-head nanoGPT campaign
• splits packed Q/K/V projections into four output-row head matrices
• splits packed O into four input-column head matrices
• runs WeightWatcher separately on all 16 attention-head matrices per checkpoint
• produces raw and clip-xmax alpha trajectories, full WeightWatcher CSV output, five-seed confidence intervals, paired optimizer differences, zoomed plots, and an HTML report
• validates completion markers, permanent checkpoint inventory, tensor shapes, finiteness, and checkpoint SHA-256 identities
• caches every checkpoint result so interrupted analysis resumes safely

Validation

• python3 -m py_compile analyze_four_head_by_head.py
• synthetic 128x128 slicing smoke test verified 16 distinct matrices: Q/K/V as 32x128 row slices and O as 128x32 column slices

This is read-only post-hoc analysis; it does not modify campaign results or require retraining.