Add RFF for simple additive kernels

[khurram-ghani]

Add RFF support for additive kernels. Re-using the existing multi-output structure as much as possible, and generalising it.

This adds RFF for simple additive kernels only. So kernels can now be Matern/RBF or multi-output or Sum. It doesn't support mixing muti-output with combination kernels for instance; that would get complex if we start considering cases where only some of the latent-kernels are a combination etc.

In the combination (i.e. Sum) kernel case, the number of feature is multiplied by the number of sub-kernels. I tried to keep them as a separate dimension (first one), as in the multi-output case, but that doesn't work well with trieste/gpflow. The main reason being that normal kernels and multi-output kernels behave slightly differently in gpflow, and adding an extra dimension to the covariance matrix causes issues in the normal kernel case. So for this gpflux PR, the dimension is kept separate in most places, but when the features are returned in call, it is flattened into the last (features) dimension.

The generalisation of the batching may be useful for any future combination kernels and batching functionality we add support for.