fix langevin corrector alpha to general num timesteps

[danielzuegner]

Addresses issue raised in #79 where the $\alpha$ coefficient for the Langevin corrector was not generalizing for different numbers of integration time steps.

The way it's generalized now works as follows:

In the score SDE paper, the scaling coefficient in the noise distribution $p(x_t | x_0)$ is defined as $\sqrt{\bar{\alpha}} = \exp\left(-0.5\int_0^t \beta(s) ds\right)$, so $\bar{\alpha} = \exp\left(-\int_0^t \beta(s) ds\right)$ (Eq. 29). In the discrete case $\bar{\alpha} = \prod_{t=1}^T \alpha_t$, i.e., the cumulative product of the per-step $\alpha_t$ coefficients. Thus, for the continuous case we obtain the equivalent by dividing the cumulative products for $t$ and $t + dt$, i.e., $\alpha_t = \frac{\bar{\alpha_t}}{\bar{\alpha}_{t + dt}}$. This happens in these lines in the code:

alpha_bar = sde._marginal_mean_coeff(t) ** 2
alpha_bar_before = sde._marginal_mean_coeff(t + dt) ** 2
alpha = alpha_bar / alpha_bar_before

where sde._marginal_mean_coeff(t) corresponds to $\sqrt{\bar{\alpha}_t}$.

Also minor fixes to tests and MP20 config.