share_comps compares and merges the wrong vectors (all backends) #334

Found during epic #324 (component-orientation investigation). Affects n_models >= 2 with share_comps=True on all three array backends, whenever a merge fires. Same root cause as the doscaling orientation issue (stored columns are not components).

Symptom

The share metric compares the wrong vectors, and the merge ties the wrong parameters.

• Metric: identify_shared_components / AMICATorchNG._identify_shared_comps / the MLX port compare de-sphered stored columns pinv(sphere) @ A[:, ci], but component i of model h is row i of the stored block (issue Verify NG init-basin vs Newton bit-parity via Fortran load_* init-matching #24 convention). On a real 2-model fit (sample EEG, 150 iterations) the |cosine| between a stored column and the true scalp map get_sensor_mixing_matrix(h)[:, i] has median 0.27 / 0.39 per model. At comp_thresh=0.95 the shipped scan merged a pair whose true maps have |cos| 0.34, while the real best matches (|cos| 0.95-0.96) scored 0.15-0.45 and never merged; with 5 models at 0.9, 30 merges reached |cos| 0.001. A planted exact duplicate component scores 0.74 and is not merged at 0.99.
• Merge: folding stored column cj into ci ties one sphered channel's loadings across the two models' components (a row of each true mixing matrix), not the two components' mixing vectors, and the gm-weighted dAk/zeta average then averages the wrong things. Seeded from a merged state through Fortran's load_comp_list, the shipped update diverges from the reference immediately (A 0.19-0.20, LL 6.4e-4 after 3 iterations); a row-layout prototype matches it to the no-merge noise floor. Forcing the shipped fold on a planted identical pair changes the LL by -0.186 (the reference fold: exactly 0).
• End to end (2 models, 300 iterations, share off LL -3.3439): shipped comp_thresh=0.95 ends at -3.4141 with Newton fallbacks; the prototype ends at -3.3416 with 3 merges of true pairs (|cos| 0.96-0.97).
• At the default comp_thresh=0.99 nothing merged on the sample EEG, so default settings on that data were unaffected; other data can differ.

Fix (epic #324 Phase 8)

• Store A with components as rows (shape (n_comps, n), comp_list indexes rows); every per-model block keeps today's n x n matrix, so every configuration without sharing stays byte-identical. Metric on pinv(sphere) @ A.T; merge ties rows; update and accessors follow. All three backends.
• Tests that pin semantics, not just cross-backend agreement: grouped sources have identical component maps; a planted duplicate is merged and LL/transform/maps are unchanged; the metric vector equals get_sensor_mixing_matrix(h)[:, i]; a gated native oracle from a merged load_comp_list state.
• Persistence: convert unmerged multi-model saves losslessly (format bump); refuse merged ones with a message to refit. The EEGLAB A export becomes the reference layout for n_models > 1.
• Changelog and a note to refit fits made with share_comps=True where merges fired.