Upper bound nnz propagation over spark instructions

Due to lazy evaluation, many spark instructions propagate nnz only if they can be exactly inferred from the input meta data. Especially for ultra-sparse matrices this poses a challenge because intermediates quickly reach a state where they're assumed to be dense. However, many operations such as right indexing allow to upper bound the nnz with the nnz in the input, which can be propagated and exploited in subsequent instructions.