Add support for external network.

Dynamically load external network from libnccl-net.so.
Add init function in networks.
Move PCI scoring to net.cu, only ask transport to provide a path.
Simplify CUDA PCI path detection.
Add dummy external network


[ROCm/rccl commit: 0d3a20f96d]

projects/rccl/src/Makefile

@@ -8,7 +8,7 @@ include ../makefiles/common.mk
 include ../makefiles/version.mk
 
 ##### src files
-INCEXPORTS  := nccl.h
+INCEXPORTS  := nccl.h nccl_net.h
 LIBSRCFILES := init.cu ring.cu bootstrap.cu transport.cu misc/group.cu \
 		misc/nvmlwrap.cu misc/ibvwrap.cu misc/rings.cu misc/utils.cu misc/enqueue.cu \
 		transport/p2p.cu transport/shm.cu transport/net.cu transport/net_socket.cu transport/net_ib.cu \
@@ -80,6 +80,11 @@ $(INCDIR)/%.h : %.h
 	mkdir -p $(INCDIR)
 	cp -f $< $@
 
+$(INCDIR)/nccl_%.h : include/nccl_%.h
+	@printf "Grabbing   %-35s > %s\n" $< $@
+	mkdir -p $(INCDIR)
+	cp -f $< $@
+
 $(OBJDIR)/%.o : %.cu
 	@printf "Compiling  %-35s > %s\n" $< $@
 	mkdir -p `dirname $@`