Support for non-zero prior mean

[st--]

We currently don't have any tests that a non-zero prior mean actually works.

• SparseVariationalApproximation: supported (with a bug fixed in SparseVariationalApproximation cleanup #86)
• LaplaceApproximation: might or might not require extra work, need to investigate (currently throws error)
• ExpectationPropagation: might or might not require extra work, need to investigate (currently throws error)

[hmray2025]

I'm trying to fit a GP using the LA and think a non-zero mean function may be helpful. Is this an issue I could help address? I would posit that three different tests are needed:

1. Training with non-zero mean, both linear and perhaps more complex?
2. Prediction can be made and output shapes are consistent.
3. Prediction values align with a set of test points.