slight improvements to IndexKernel for the rank == 0 (diagonal) case

[rmgarnett]

This PR tweaks IndexKernel to have slightly better performance/behavior in the corner case of rank == 0, where the underlying covariance matrix is diagonal. Although this case is atypical from the point of view of multitask GPs, I have found this construction useful as a component of larger covariance structures.

Two changes are proposed:

• the kernel only registers the covar_factor parameter (the low-rank component when rank > 0) when rank > 0
• in the rank == 0 case, the covar_matrix() method returns a DiagLinearOpeator rather than a PsdSumLinearOperator(RootLinearOperator, DiagLinearOperator)

[Balandat]

In principle it seems reasonable enough to me to support this.

However, looking at forward and _eval_covar_matrix the code as is will cause some issues. Namely in _eval_covar_matrix (called from forward) we will get an AttributeError when we try this from self.covar_factor. Which means that we should also add a unit test case to cover the rank==0 behavior.

Taking a step back, if we handled that correctly, we probably shouldn't return an InterpolatedLinearOperator in forward - or at least we'd want to special case InterpolatedLinearOperator to handle a diagonal base_linear_op (it doesn't seem that this is currently the case).

[rmgarnett]

I've added a branch to _eval_covar_matrix() to handle the diagonal case to partially address @Balandat's comments.

[gpleiss]

LGTM, but I'll let @Balandat make the final approval.

[Balandat]

lgtm. I think at least some of the special casing of using a DiagLinearOperator is actually happening inside InterpolatedLinearOperator so we'll probably get some gains there already.

[Balandat]

I guess I would still like to see this covered by a unit test though :)