Refactor KalmanFilter class as OpFromGraph

[jessegrabowski]

The most requested feature for the statespace module is to handle multiple time series in a single MCMC run. This will require support for batch dimensions. I think the easiest way to attack this will be to refactor the KalmanFilter class to be an OpFromGraph. My original class-based design was inspired by the statsmodels implementation, but it doesn't take full advantage of pytensor.

An additional advantage of this will be the ability to define a custom gradient. See #332.

Finally, it will let us handle special case filters via rewrites, rather than asking the user to pick a filter up front.