Raptor-T: A Fused and Memory-Efficient Sparse Transformer for Long and Variable-Length Sequences | Synapse