Add Alltoallv RCCL kernel implementation (#269)

* Add alltoallv API and implementation

* Extend Rome P2P channel limit to multinode and alltoall kernels

* topo_expl: fix compilation and sync up with main

* gtest: use RCCL alltoallv API

* Code review changes

test/test_AllToAllv.cpp

@@ -35,29 +35,10 @@ namespace CorrectnessTests
                 sdispls[i] = rdispls[i] = disp;
                 disp += scount;
             }
-            for (int i = 0; i < numDevices; i++) {
-                if (sendcounts[i] != 0) {
-                  ncclSend(
-                      ((char*)dataset.inputs[r]) + sdispls[i] * DataTypeToBytes(dataType),
-                      sendcounts[i],
-                      dataType,
-                      i,
-                      comms[r],
-                      streams[r]);
-                }
-                if (recvcounts[i] != 0) {
-                  ncclRecv(
-                      ((char*)dataset.outputs[r]) + rdispls[i] * DataTypeToBytes(dataType),
-                      recvcounts[i],
-                      dataType,
-                      i,
-                      comms[r],
-                      streams[r]);
-                }
-            }
+            ncclAllToAllv((char*)dataset.inputs[r], sendcounts, sdispls,
+              (char*)dataset.outputs[r], recvcounts, rdispls, dataType, comms[r], streams[r]);
         }
         ncclGroupEnd();
-
         // Wait for reduction to complete
         Synchronize();
 
@@ -89,6 +70,6 @@ namespace CorrectnessTests
                                 testing::Values(2,3,4,5,6,7,8),
                                 // In-place or not
                                 testing::Values(false),
-                                testing::Values("")),
+                                testing::Values("RCCL_ALLTOALL_KERNEL_DISABLE=0", "RCCL_ALLTOALL_KERNEL_DISABLE=1")),
                             CorrectnessTest::PrintToStringParamName());
 } // namespace