[coll] Implement shutdown for tracker and comm. (#10208)

- Force shutdown the tracker.
- Implement shutdown notice for error handling thread in comm.

rabit/include/rabit/internal/socket.h

@@ -100,6 +100,24 @@ std::enable_if_t<std::is_integral_v<E>, xgboost::collective::Result> PollError(E
   if ((revents & POLLNVAL) != 0) {
     return xgboost::system::FailWithCode("Invalid polling request.");
   }
+  if ((revents & POLLHUP) != 0) {
+    // Excerpt from the Linux manual:
+    //
+    // Note that when reading from a channel such as a pipe or a stream socket, this event
+    // merely indicates that the peer closed its end of the channel.Subsequent reads from
+    // the channel will return 0 (end of file) only after all outstanding data in the
+    // channel has been consumed.
+    //
+    // We don't usually have a barrier for exiting workers, it's normal to have one end
+    // exit while the other still reading data.
+    return xgboost::collective::Success();
+  }
+#if defined(POLLRDHUP)
+  // Linux only flag
+  if ((revents & POLLRDHUP) != 0) {
+    return xgboost::system::FailWithCode("Poll hung up on the other end.");
+  }
+#endif  // defined(POLLRDHUP)
   return xgboost::collective::Success();
 }
 
@@ -179,9 +197,11 @@ struct PollHelper {
     }
     std::int32_t ret = PollImpl(fdset.data(), fdset.size(), timeout);
     if (ret == 0) {
-      return xgboost::collective::Fail("Poll timeout.", std::make_error_code(std::errc::timed_out));
+      return xgboost::collective::Fail(
+          "Poll timeout:" + std::to_string(timeout.count()) + " seconds.",
+          std::make_error_code(std::errc::timed_out));
     } else if (ret < 0) {
-      return xgboost::system::FailWithCode("Poll failed.");
+      return xgboost::system::FailWithCode("Poll failed, nfds:" + std::to_string(fdset.size()));
     }
 
     for (auto& pfd : fdset) {