feat: expose convergence diagnostics from VBPCA.fit() and select_n_components()

[jc-macdonald]

Closes #99

Summary

Expose convergence diagnostics that pca_full() already computes but VBPCA.fit() previously discarded. This enables downstream analysis of convergence behavior — particularly important for optimizing convergence parameters where wall_seconds is an unreliable proxy.

Changes

_converge.py

• New _cost_criteria_tagged() returns (reason_tag, message) tuples instead of bare strings
• convergence_check() stores a structured _convergence_reason tag in the learning curve dict when a criterion fires
• Reason tags: angle, earlystop, rms_plateau, cost_plateau, cfstop_rel, cfstop_curv, composite, slowing_down

_pca_full.py

• After the training loop, promotes _convergence_reason → convergence_reason in lc, defaulting to "maxiters"

estimators.py

Three new fitted attributes on VBPCA:

• n_iter_: int | None — iterations completed
• convergence_reason_: str | None — structured reason tag or "maxiters"
• learning_curve_: dict | None — full per-iteration history (rms, prms, cost, angle, phase timings)

model_selection.py

• _fit_candidate() trace entries now include n_iter and convergence_reason

Tests

• 4 new tests in test_estimators.py: diagnostics exposed after fit, maxiters reason, angle reason, None before fit
• 1 new test in test_model_selection.py: trace entries contain convergence diagnostics

Motivation

Investigation revealed that niter_broadprior=100 (the default) suppresses convergence checks, causing 30–93% iteration waste with negligible quality improvement:

Case	bp=100	bp=0	Waste	Quality Δ
small (10×20)	110	68	+62%	negligible
medium (50×100)	175	112	+56%	0.002 RMS
medium+missing	220	114	+93%	0.001 RMSE
large (100×200)	147	114	+29%	—

These diagnostics enable data-driven optimization of niter_broadprior and convergence thresholds.