docs: notebook showing different step size schedules

[jeremiecoullon]

1. Add cyclical sgmcmc step size schedule to diffusions.py:

def cyclical_schedule(alpha_0, M, K):
    def schedule(i):
        mod_term = (i-1) % jnp.ceil(K/M)
        return alpha_0*0.5*(jnp.cos( jnp.pi*mod_term /jnp.ceil(K/M) ) + 1)
    return schedule

Usage:

dt_cyclical = cyclical_schedule(1e-3, 5, 1000)
my_sampler = build_sgld_sampler(dt_cyclical, loglikelihood, logprior, data, batch_size)

2. Then add a notebook showing the different schedules: constant, Welling & Teh, cyclical. Perhaps do OOD testing for MNIST (like in the cyclical paper linked above)