Rewrite StandardScaler in terms of equistore operations

[Luthaf]

StandardScaler should be implemented in term of equistore.mean, equistore.std, … Right now, it contains a manual re-implementation of these operations.

See #15 (comment)

[agoscinski]

Maybe it is actually useful to use the mean_over_samples and variance_over_samples function here.

Originally posted by @PicoCentauri in #15 (comment)

[PicoCentauri]

I was looking into this and since we are using a weighted mean we first have to add an equistore.average_over_samples function with a weight option - similar to numpy.average. Currently we only have a classical mean. Pinging also @DavideTisi here :-)

[agoscinski]

I think we should standardize the features and gradients more consistently. At the moment in the standardization we divide by two different terms. One we get from the standard deviation of the values and the other from the standard deviation of the gradients, but we should use both times the standard deviation of the values to be consistent. (My bad, I implemented this)

$$\hat{a} = \frac{a}{\sigma}$$

$$ \frac{\partial\hat{a}}{\partial x} = \frac{\partial a}{\partial x} \frac1{\sigma} $$