Open
Description
We have an implementation of a square root filter which should be numerically more stable than the standard kalman filter, and potentially a bit faster (citation needed).
I had removed it from tests and options because it involves a QR decomposition, which at the time didn't have gradients defined. So it wasn't super useful. In the meantime, though, we got QR gradients. So we should bring back tests for this, check if it's working, and if so, bring it back in the user facing API.