Model and mapper for distributed buffer (only unshared tensors)

[gilbertmike]

• Support for network topologies (mesh and all-to-all using switch supported).
• Model and mapper support for distributed buffers, but only for unshared tensors. (Shared tensors in future PR.)
  • Note on flattening: when an Array node with parallel nodes (inside the nodes key) exists, the parallel nodes are turned into "logical" nodes by removing their fanout during flattening. The "logical" node represents the combined capacity of the distributed buffers. This conversion allows the mapping to occur as usual, but the cost will account for the fanout of spatial nodes underneath the storage node of logical nodes.
• Update for network model to use the new throughput interface.
• Throughput interface behavior updated:
  • If only latency is specified, a warning is emitted and auto-conversion is done.
  • If throughput is specified, the user is assumed to already be using new interface, no warning or conversion done (used to raise error).