[Triton-MLIR][BACKEND] Fix the membar pass to add missing barriers caused by scf.for (#933)

1. Add missing barriers and revert the previous temporary solution
2. Extract the `run` method from membar analysis because the membar
analysis should have two phases, including construction, which doesn't
modify any IR, and modification, which adds barrier IRs. Hope this could
make the use of membar clear.

include/triton/Analysis/Membar.h

@@ -29,7 +29,11 @@ public:
   /// The following circumstances are not considered yet:
   /// - Double buffers
   /// - N buffers
-  MembarAnalysis(Allocation *allocation) : allocation(allocation) { run(); }
+  MembarAnalysis(Allocation *allocation) : allocation(allocation) {}
+
+  /// Runs the membar analysis to the given operation, inserts a barrier if
+  /// necessary.
+  void run();
 
 private:
   struct RegionInfo {
@@ -82,10 +86,6 @@ private:
     }
   };
 
-  /// Runs the membar analysis to the given operation, inserts a barrier if
-  /// necessary.
-  void run();
-
   /// Applies the barrier analysis based on the SCF dialect, in which each
   /// region has a single basic block only.
   /// Example: