feat: Variational expectations.

[daniel-dodd]

Feature Request

GPJax's SVGP computes variational expectations $\int \text{d} q(f(x_i)) \log( p(y_i|f(x_i))$ via Gauss-Hermite quadrature.

This is superfluous in cases e.g., of a Gaussian likelihood when the integrand is known is closed form. It would be beneficial to override this automatically.

Additionally, it would nice to offer alternatives such as Monte Carlo.

[daniel-dodd]

Questions:

1. When the integrand is known in closed form, where should we put it?
2. Could we define the integration strategy when initialising StochasticVI?

[daniel-dodd]

@thomaspinder Resolving the Gaussian case for the analytical variational expectation, would allow us to test the collapsed bound agrees with Natural gradients after one gradient update with fixed hyper-parameters (#90). Currently, both bounds almost agree, but differ due to the quadrature approximation.

[thomaspinder]

Closing in lieu of #270