Sample from constrained kernel params before model fitting

[henrymoss]

This PR allows our kernel initialization routine for GPR models to sample kernel parameters with (Sigmoid) constraints. Previously, this routine only sampled parameters with hyper-priors.

Our sampling strategy follows the following hierarchy

1. If a param is constrained with a sigmoid bijector then we sample from its constrained domain (ignoring its priors if it has them).
2. If a param is not constrained but has a prior then we sample from the prior
3. otherwise we do not sample the param and fix its values to its initialization.

This new change will be detailed in my imminent notebook overhaul.

[henrymoss]

To use this in practice, do something like
upper = (search_space.upper-search_space.lower)*tf.math.sqrt(tf.cast(len(search_space.upper),dtype=tf.float64))
lower = upper / 100

model.kernel.lengthscales = gpflow.Parameter(
        model.kernel.lengthscales, transform=tfp.bijectors.Sigmoid(low=lower, high=upper)
    )

when defining your model