Implement memory-efficient subsampling of non-square or non-sparse relation data

[einbandi]

So far, subsampling is only properly implemented for dense, square tensors and arrays. Other RelationData subclasses internally are converted to dense representations for subsampling. This is memory-inefficient. For the triangular representations, the index lists must be converted accordingly to enable a direct indexing without having to call to_square...(). For the sparse matrix representation it should be enough to check that subsampling works properly with a similar implementation as for the dense case.