Pre-compute loss function in `finite_horizon`

[LarrySnyder]

Tried working on this, but it didn't seem to speed things up much. Why?

Added this code right before main loop:

    # Pre-calculate standard normal loss function values.
    min_z, max_z, step_z = -4, 4, 0.01
    loss_table = lf.standard_normal_loss_dict(start=min_z, stop=max_z, step=step_z)
    comp_table = lf.standard_normal_loss_dict(start=min_z, stop=max_z, step=step_z, complementary=True)

And this code when n(y) and \bar{n}(y) are calculated:

                if [...some option is set...]:
                    z = (y - demand_mean[t]) / demand_sd[t]
                    if z < min_z:
                        n = demand_mean[t] - y
                        n_bar = 0
                    elif z > max_z:
                        n = 0
                        n_bar = y - demand_mean[t]
                    else:
                        n = nearest_dict_value(z, loss_table) * demand_sd[t]
                        n_bar = nearest_dict_value(z, comp_table) * demand_sd[t]
                else:
                    n, n_bar = lf.normal_loss(y, demand_mean[t], demand_sd[t])

Maybe nearest_dict_value is too slow?