_return_predictions should always run the Tables.materializer

[rafaqz]

For most predict return types Tables.materializer is called. But the AbstractVector method just returns a vector: https://github.com/JuliaStats/StatsModels.jl/blob/master/src/statsmodel.jl#L149-L154
For some objects it would be preferable if it called the materializer with T((; predictions = out)).

For example, a DimensionalData.jl/Rasters.jl stack or array can be used as a table, but it cant get a stack/array back without Tables.materializer being called. This means predict on multimensional spatial variables returns non-spatial single dimension predictions that have to be manually reshaped, instead of it being automatic.

This will probably be a breaking change.