Combine backtrace print into one string to avoid interleaving. (#56961)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/56961

As described in https://github.com/pytorch/pytorch/issues/56583, the
backtrace amongst several processes was garbled.
https://github.com/pytorch/pytorch/pull/56198 would've alleviated this to some
extent, but this PR combines all the logging into just one string to reduce
interleaving further.
ghstack-source-id: 128730047

Test Plan: waitforbuildbot

Reviewed By: cbalioglu

Differential Revision: D28013191

fbshipit-source-id: 8bd8978a92ee2fbcd18472e1293d5809455b411b

BUILD.bazel

@@ -83,6 +83,7 @@ cc_library(
     copts = ["-DCAFFE2_BUILD_MAIN_LIB"],
     deps = [
         ":c10_headers",
+        "@fmt",
     ] + if_cuda(
         ["@cuda"],
         [],

c10/CMakeLists.txt

@@ -59,6 +59,7 @@ endif()
 if(${USE_GLOG})
     target_link_libraries(c10 PUBLIC glog::glog)
 endif()
+target_link_libraries(c10 PRIVATE fmt::fmt-header-only)
 
 find_package(Backtrace)
 if(Backtrace_FOUND)

c10/util/signal_handler.cpp

@@ -1,5 +1,6 @@
 #include <c10/util/Backtrace.h>
 #include <c10/util/signal_handler.h>
+#include <fmt/format.h>
 
 #if defined(C10_SUPPORTS_SIGNAL_HANDLER)
 
@@ -162,10 +163,14 @@ void FatalSignalHandler::stacktraceSignalHandler(bool needsLock) {
     pthread_mutex_lock(&writingMutex);
   }
   pid_t tid = syscall(SYS_gettid);
-  std::cerr << fatalSignalName << "(" << fatalSignum << "), PID: " << ::getpid()
-            << ", Thread " << tid << ": " << std::endl;
-  std::cerr << c10::get_backtrace();
-  std::cerr << std::endl;
+  std::string backtrace = fmt::format(
+      "{}({}), PID: {}, Thread {}: \n {}",
+      fatalSignalName,
+      fatalSignum,
+      ::getpid(),
+      tid,
+      c10::get_backtrace());
+  std::cerr << backtrace << std::endl;
   if (needsLock) {
     pthread_mutex_unlock(&writingMutex);
     pthread_cond_signal(&writingCond);