Implement local map #4
Description
Currently, only the global map case is supported (i.e. the relabeling map is shrunk up to fit into each single GPU); local map is the case when the local maps in the different GPUs are not overlapped, these can be shared with ring communication and independently applied in the round to the local edges.
This approach is more scalable but seems unlikely to happen (the map has the size of the rows (not nnz)). Currently, the limitation for global map is the MAX_INT32, not the GPU memory.