Super duper thompson sampler

[henrymoss]

We can now do continuous Thompson sampling with Trieste!

By continuous, I mean optimizing sample draws (trajectories) using L-BFG-B rather than just choosing the best points over a discrete set (as currently performed by our DiscreteThompsonSamplingRule).

In order to avoid retracing I added some update/resample logic to our RFF trajectory sampler, inspired by our AcquisitionFunctionBuilders. I also changed the expected input shape of trajectories so that they behave more like acquisition functions and so can be optimized using our optimizers.

In this PR, we sequentially maximize trajectory samples using our GreedyAcquisitionFuncton logic.

Future PRs will add (in order)

1. A Matheron's rule trajectory sampler for our VGP and SVGP models to allow continuous Thompson sampling with these models.
2. Added support for parallel continuous Thompson sampling, i.e maximising all the trajectories in parallel using our newish parallel acquisition optimiser.

[uri-granta]

Looks great! Just some minor comments.