Reduce linking time through more parallel jobs (#657)

CMakeLists.txt

@@ -442,8 +442,8 @@ if("${HIP_COMPILER}" MATCHES "clang")
   find_program( hipcc_executable hipcc )
   execute_process(COMMAND bash "-c" "${hipcc_executable} -help | grep 'parallel-jobs'" OUTPUT_VARIABLE hipcc_parallel_jobs)
   if("${hipcc_parallel_jobs}" MATCHES "parallel-jobs")
-    target_compile_options(rccl PRIVATE -parallel-jobs=8 PRIVATE -Wno-format-nonliteral)
-    target_link_libraries(rccl PRIVATE -parallel-jobs=8)
+    target_compile_options(rccl PRIVATE -parallel-jobs=12 PRIVATE -Wno-format-nonliteral)
+    target_link_libraries(rccl PRIVATE -parallel-jobs=12)
   endif()
 
   # RCCL static lib uses -fgpu-rdc which requires hipcc as the linker and archiver