Use only LDA exchange in SAD guess instead of SVWN5.

[susilehtola]

In our study on initial guesses in J. Chem. Phys. 152, 144105 (2020), we examined the importance of including correlation in the guess potential. However, we found out that the correlation functional only plays a minor importance, and that it also can sometimes lead to unexpected results; as far as I remember, the correlated potentials were sometimes more shallow than those obtained from an exchange only calculation.

This PR removes the correlation contribution from the initial guess.

[ilfreddy]

I guess the failed tests are a consequence of the different starting guess, given that such tests are not run to convergence, but only for 1-2 SCF cycles. Therefore the reference results should most likely be updated in this case.
However, there is a lot happening with MRChem at present (many new features, several changes in the code) and I would therefore like to limit the number of PR where reference results must be updated as this can potentially introduce new errors.
We discussed a policy where PRs that require changes in the reference results should also be subject to a more thorough scrutiny (a test suite of converged results at high precision) to make sure that changing the reference results is safe.
In conclusion this PR is sensible, but it should be postponed until such a mechanism is in place, in my opinion.