Cheaper threadfence for gfx942 in postPeer [1/N]: enable for single node allreduce (#1766)

Boosts single node bfloat16 allreduce performance by up to 20% for some data sizes and provides gating with the RCCL_GFX942_CHEAP_FENCE_OFF environment variable

src/include/comm.h

@@ -488,6 +488,7 @@ struct ncclComm {
   int node;
   int nNodes;
   int rcclUseOneSlice; // RCCL: true if this comm is using one slice per primitive
+  int gfx942CheapFenceOff; // RCCL: true if gfx942 cheap fence is disabled
   int localRank;
   int localRanks;
   int maxLocalRanks;

src/include/device.h

@@ -305,7 +305,7 @@ struct alignas(16) ncclDevWorkColl {
   //   nChannels == (channelHi - channelLo) + 1
   uint32_t channelLo:8, channelHi:8;
   uint32_t nWarps:8;
-  uint32_t redOpArgIsPtr:1, regUsed:1, netRegUsed:1, oneNode:1, direct:2, isOneRPN:1, rcclUseOneSlice:1;
+  uint32_t redOpArgIsPtr:1, regUsed:1, netRegUsed:1, oneNode:1, direct:2, isOneRPN:1, rcclUseOneSlice:1, gfx942CheapFenceOff:1;
   uint32_t root:30, connIndex:2;
   uint16_t pivotA2ANumBiRings:15, profilerEnabled:1;
   void* recvbuff;