Cosine Annealing with Warm Restarts

SGDR decays the learning rate along a cosine curve, then jumps back up. Cycle $i$ lasts $T_i = T_0 \cdot m^i$ steps.

Within a cycle, at position $t_{cur}$: $\eta = \eta_{min} + \tfrac12(\eta_{max} - \eta_{min})(1 + \cos(\pi\,t_{cur}/T_i))$.

Implement sgdr_lr(step, lr_max, lr_min, T0, mult=1) for integer step >= 0.