Simplify EKI update, correct dims error in nondeterministic settings

[ilopezgp]

This PR simplifies substantially the update of the EKI algorithm.

• The computation of the cross-covariance matrices was done through a loop, which is unnecessary since we are already loading the Statistics module.
• Everything was transposed for some reason. Now all data are stored in columns, similarly to the rest of the code.

The new construction produces the same results (tested with an assert before this PR). Note that I have not added the Bessel correction to the cross-covariances cov_ug and cov_gg to ensure consistency with the papers and the previous version. We can discuss this.

In addition, I have fixed an issue with the construction of y, which was Array{FT,1} for !deterministic_forward_map.

[odunbar]

LGTM! Thanks.

As for responses:

1. i didn't realize that the cov(u,g,...) existed when first implementing this (i almost wouldn't be surprised if it didn't!)
2. The transpose was there because initially there was a clash in different parts of the code then we unified everything, and the simplest fix at the time was to just permute things going into this function.
3. With the y construction, i guess it wasn't an issue before because of broadcasting? either way this is fine with me

[ilopezgp]

Yes, I think the cov functionality for cross-covariances might be a new implementation. And I agree that this was difficult to do when unifying all notation everywhere, which already required a lot of work.

[ilopezgp]

bors r+

[bors]

Build succeeded:

• docs-build
• test_success