Fix build without roctx

With ROCTX=OFF (or roctracer not found) rccl defines NVTX_NO_IMPL and core.h
includes nvtx_stub.h, but several sources include nvtx.h directly and rely on
headers that only roctx.h pulls in.
diff --git a/src/algorithms/dda/ipc/ipc_init.cu b/src/algorithms/dda/ipc/ipc_init.cu
index 73faafd..ce7bee5 100644
--- a/src/algorithms/dda/ipc/ipc_init.cu
+++ b/src/algorithms/dda/ipc/ipc_init.cu
@@ -13,6 +13,7 @@
 #include "debug.h"
 #include "algorithms/dda/dda_init_detail.h"
 #include "algorithms/dda/ipc/ipc_mem_handler.h"
+#include <iostream>
 
 #include <cuda_runtime.h>
 
diff --git a/src/include/nvtx.h b/src/include/nvtx.h
index 16e815f..8b733a4 100644
--- a/src/include/nvtx.h
+++ b/src/include/nvtx.h
@@ -8,6 +8,10 @@
 #ifndef NCCL_NVTX_H_
 #define NCCL_NVTX_H_
 
+#ifdef NVTX_NO_IMPL
+#include "nvtx_stub.h"
+#else
+
 #include "nvtx3/nvtx3.hpp"
 #include "roctx.h"
 
@@ -243,4 +247,6 @@ private:
 
 extern void initNvtxRegisteredEnums();
 
+#endif // NVTX_NO_IMPL
+
 #endif
diff --git a/src/include/nvtx_stub.h b/src/include/nvtx_stub.h
index d393371..23a989b 100644
--- a/src/include/nvtx_stub.h
+++ b/src/include/nvtx_stub.h
@@ -14,6 +14,7 @@ struct nccl_domain {
 };
 
 #define NVTX3_FUNC_RANGE_IN(domain)
+#define NCCL_NVTX3_FUNC_RANGE
 #define nvtxNameOsThreadA(syscall, thread)
 #define NVTX3_FUNC_WITH_PARAMS(N, T, P)
 #define NVTX3_PAYLOAD(...) __VA_ARGS__
diff --git a/src/include/rccl_common.h b/src/include/rccl_common.h
index 5cca28d..a276a55 100644
--- a/src/include/rccl_common.h
+++ b/src/include/rccl_common.h
@@ -22,6 +22,7 @@ THE SOFTWARE.
 #ifndef RCCL_COMMON_H_
 #define RCCL_COMMON_H_
 #include "nccl_common.h"
+#include "nccl_tuner.h"
 #include "nccl.h"
 #include "param.h"
 #include "core.h"
diff --git a/src/transport/net.cc b/src/transport/net.cc
index 9b08181..7f7ba6c 100644
--- a/src/transport/net.cc
+++ b/src/transport/net.cc
@@ -21,6 +21,7 @@
 #include "shm.h"
 #include "compiler.h"
 #include <atomic>
+#include <map>
 #include <assert.h>
 #include "graph.h"
 #include "graph/topo.h"