Use hipGraph instead of cudaGraph (#487)

src/include/comm.h

@@ -15,8 +15,6 @@
 // [/RCCL]
 
 #if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
-  typedef void *cudaGraph_t;
-  typedef void *cudaGraphNode_t;
   #define HIPRT_CB
 #else
 #include "collectives.h"
@@ -216,7 +214,7 @@ struct ncclComm {
   struct ncclQueueInfo* enqueueInfo;
   int nQueueInfoCreated;
   int nQueueInfoDestroyed;
-  cudaGraphNode_t lastSetupNode;
+  hipGraphNode_t lastSetupNode;
   unsigned long long lastCudaGraphId;
   int driverVersion;
   pfn_cuMemGetAddressRange_t pfnCuMemGetAddressRange;

src/include/enqueue.h

@@ -28,8 +28,8 @@ ncclResult_t ncclSetupP2pKernel(struct ncclInfo* info);
 ncclResult_t ncclSetupAsyncKernels(struct ncclComm* comm);
 template<int USING_CUDA_GRAPH>
 void HIPRT_CB ncclEnqueueHostSetup(void* arg);
-ncclResult_t ncclGetCudaGraph(ncclComm_t comm, cudaGraph_t* graph);
-ncclResult_t ncclCudaGraphHostSetup(ncclComm_t comm, cudaGraph_t graph);
+ncclResult_t ncclGetCudaGraph(ncclComm_t comm, hipGraph_t* graph);
+ncclResult_t ncclCudaGraphHostSetup(ncclComm_t comm, hipGraph_t graph);
 
 struct ncclBuffRegInfo {
   void* sendbuffsBase[NCCL_MAX_INTRA_RANKS];