Fix erroneous posterior covariance with preconditioned CGD

[SebastianCallh]

We expect the P-CGD strategy to converge to the same posterior as vanilla CGD, but currently the variance doesn't seem to be updated properly - the posterior variance is much too big.

[sethaxen]

I'm happy to look into this.

[sethaxen]

Seems to relate #3 (comment)

[SebastianCallh]

Thank you, that's super appreciated!

Let me know if I can help somehow or if the current implementation needs clarifications.

[SebastianCallh]

I gave this some more time and reproduced the issue in a self-contained file with preconditioned conjugate gradients for regular GPs. Didn't make any more progress yet but I put the code in this gist
https://gist.github.com/SebastianCallh/738fda183bf8675697ec954e160c75ea