Varying slopes: design operator applies per-observation loadings + term-major coefficient offsets

[schroedk]

Implements #71.

• FactorMeta → TermMeta (pure-rename commit first): terms carry an intercept flag plus slope column indices into the frame's continuous columns; Design::new lowers slope-bearing Effects, laid out term-major (offset[t] + c·L_t + level, columns [intercept?, slopes…])
• Gather: one fused sweep per term via const-generic gather_term (hot shapes dispatch to it, rare shapes take a fused dynamic fallback); scatter: const-generic multi-column kernels over the existing strategy dispatch — C=1 compiles to the previous single-column kernels. Kernel shapes chosen by A/B measurement (1M/5M/20M; arm unification verified tie at 100K–40M). Each direction's dispatch lives with its kernels (gather.rs / scatter.rs)
• Plain designs stay bit-identical; no n=1M criterion regression vs saved baseline
• Slope designs build and the operator is fully exercisable; Solver::new rejects solving them (SlopesNotYetSupported) until the transform slices (Varying slopes: single slope on one factor (V=1) — centering + back-transform #59+)