Pure pytensor GP

[jessegrabowski]

Very early draft, do not look

TODOs (subject to evolution):

• [ ] Marginalization of GP random variables using the MarginalModel API
• [ ] New model transformation for creating new random variables by conditioning on data
• [ ] More covariance functions
• [ ] Covariance function specific rewrites (optimizations when adding/multiplying?)
• [ ] Rewrite of covariance to spectral density when using HSGP random variable
• [ ] Rewrite to approximations via pm.fit