Handle random_seed setting at higher level in package

[edwardchalstrey1]

Should close #465

Changes

• Make set_random_seed function available to both AutoEmulate class and Emulator class (and everything inheriting from it) via ConversionMixin
• I removed setting random_seed in the models directly, but then I noticed with was needed for RF, so restored it there as the random_state is a required parameter in the actual sklearn to ensure it behaves deterministically
• I also kept the random_seed parameter in GP (and we might want to keep for all models), see my questions below

Reviewers

A few decisions need to be made here that I'd like input on:

1. I added a test for the GP model, which if we want to keep, I will set up for the other models too. The test attempts to change the random_seed and then create a new model which should have different predictions. At the moment, this fails. My question is, do we care? And assuming that we do, can you spot where I might be going wrong?
2. Might we want to refactor this into a new random_seed into a new mixin or rename this ConversionMixin? Assuming we want it to be available to AutoEmulate and Emulator, which we might not

[sgreenbury]

It might be the case that:

• we expect these to be close for different fits of the GP (perhaps it would be worth also testing the hyperparameters of the fitted model for similarity?)
• the fitting is not stochastic for this model (and others in this PR too) and has no dependence on seed

I think it might be worth documenting for each model in the init docstring the expected effect of providing a random seed.