[rabit_bootstrap_cache ] failed xgb worker recover from other workers (#4808)

* Better recovery support.  Restarting only the failed workers.

src/data/data.cc

@@ -229,7 +229,8 @@ DMatrix* DMatrix::Load(const std::string& uri,
   /* sync up number of features after matrix loaded.
    * partitioned data will fail the train/val validation check
    * since partitioned data not knowing the real number of features. */
-  rabit::Allreduce<rabit::op::Max>(&dmat->Info().num_col_, 1);
+  rabit::Allreduce<rabit::op::Max>(&dmat->Info().num_col_, 1, nullptr,
+    nullptr, fname.c_str());
   // backward compatiblity code.
   if (!load_row_split) {
     MetaInfo& info = dmat->Info();