Support for `emp_start` to apply to other non-spatial covariance structures

**Summary** The current function `emp_start` works well to speed up `us` covariance estimation. I would like to extend this to the other non-spatial covariance structures so that we start with a close approximation before going into a general optimization loop. **Additional Information** Essentially, the existing `emp_start` need only use an additional helper function that calculates a reasonable starting value for non-`us` covariance structures from the empirical covariance matrix.