fix(rdma): avoid lttng templates under extern "C"

On builds where lttng is enabled, the inclusion of nccl_ofi_tracepoint.h
pulls in <lttng/ust-utils.h>, which defines templates when __cplusplus
is defined without restraint. Those templates cannot have C linkage.

This file does not really need nccl_ofi_tracepoint.h; it only needs
config.h to test for whether NVTX tracing is enabled, and then it needs
nvtx3 headers in the case that it is so that it may define storage for
the nvtx domain handles.

Signed-off-by: Nicholas Sielicki <nslick@amazon.com>

include/nccl_ofi_config_bottom.h

@@ -5,6 +5,8 @@
 #ifndef NCCL_OFI_CONFIG_BOTTOM_H
 #define NCCL_OFI_CONFIG_BOTTOM_H
 
+#define NCCL_OFI_N_NVTX_DOMAIN_PER_COMM 8
+
 /* configure aborts if __buildin_expect() isn't available */
 #define OFI_LIKELY(x)   __builtin_expect((x), 1)
 #define OFI_UNLIKELY(x) __builtin_expect((x), 0)

include/nccl_ofi_rdma.h

@@ -9,6 +9,8 @@
 extern "C" {
 #endif
 
+#include "config.h"
+
 #include <rdma/fabric.h>
 
 #include "nccl_ofi.h"
@@ -19,8 +21,10 @@ extern "C" {
 #include "nccl_ofi_deque.h"
 #include "nccl_ofi_freelist.h"
 #include "nccl_ofi_idpool.h"
-#include "nccl_ofi_tracepoint.h"
 #include "nccl_ofi_ep_addr_list.h"
+#if HAVE_NVTX_TRACING
+#include <nvtx3/nvToolsExt.h>
+#endif
 
 /* Maximum number of rails supported. This defines the size of
  * messages exchanged during connection establishment (linear

include/tracing_impl/nvtx.h

@@ -8,8 +8,6 @@
 #if HAVE_NVTX_TRACING
 #include <nvtx3/nvToolsExt.h>
 
-#define NCCL_OFI_N_NVTX_DOMAIN_PER_COMM 8
-
 static inline void nvtx_mark_domain(nvtxDomainHandle_t domain, const char* name, uint32_t color)
 {
 	nvtxEventAttributes_t eventAttrib = {};