Implement gradient descent and verify all four rates

Implement gradient descent and empirically verify each of the four theorems:

1. Convex smooth: $O(1/t)$ on $f(x) = \sum_i\log(1+e^{-y_ia_i^\top x})$ (separable data removed so a minimiser exists — add a tiny ridge).
2. Strongly convex smooth: linear rate on a quadratic, and check the measured contraction factor against $1-1/\kappa$ and against $1-2/(\kappa+1)$.
3. Non-convex smooth: $\min_{s\le t}\|\nabla f\|$ decaying like $t^{-1/2}$ on a non-convex objective.
4. Non-smooth convex: the last iterate oscillating and the average converging on $f(x)=\|x\|_1$.