Phases 86-88: in-place eigendecomp simplification, code quality
improvements, analytical dev2 for all n_cvt, fused Uab compute
for reduced peak memory in NumPy runners.

Lazy eigendecomposition (phase 89) added and reverted — dstedc
workspace (3N²) makes it unviable at 100k+ samples.