[GraphBolt] Compress type_per_edge array so that instead of COO, we use CSC.

Our sampling code assumes that edges are sorted with respect to the edge types. Thus, we can actually use CSC to store it instead of COO. If the type per edge array for 2 vertices looks like this: [[1 1 1] [0 0 2 2 2 2]] with first vertex having 3 edges, second having 6, then this could be compressed into 3 arrays as:
`indptr=[0, 1, 3]` offsetting into: `tindptr=[0, 3, 5, 9]` and `types=[1, 0, 2]` giving actual type ids.

For example:
```python
for node_id in nodes:
for i in range(indptr[node_id], indptr[node_id + 1]):
type = types[i]
# indices values of edges with same type
indices[tindptr[i]: tindptr[i + 1]]
```