Make cov and cor similar to mean and var by removing keyword arguments.

[andreasnoack]

This fixes #13081 and the type instability that motivated that issue.

I think it is actually quite rare that users would like to provide a precalculated mean and when they do, they would like maximum performance which they wouldn't get from a keyword argument function anyway. Since precalculated means are rare, I have also chosen not to export covm and corm.

It is probably more likely that someone would like the uncorrected version of the covariance, but it is not hard to get without the keyword argument.

The dimension keyword is simply too inconsistent with what we are doing for mean, var, sum etc.

I've also moved the documentation to the method definitions and split it out on the individual methods.

[andreasnoack]

By the way, when updating the tests to the new signatures, I realized that cov and cor are not really that extensively tested for numerical accuracy. The tests are mainly testing that all the dispatch is working.