[BACKEND] Make flash attention forward pass work (#928)

This also simplifies BroadcastOp codegen

lib/Dialect/TritonGPU/Transforms/Prefetch.cpp

@@ -130,6 +130,11 @@ LogicalResult Prefetcher::initialize() {
 
   if (dotsInFor.empty())
     return failure();
+  
+  // TODO: segfault (original for still has uses)
+  // when used in flash attention that has 2 dots in the loop
+  if(dotsInFor.size() > 1)
+    return failure();
 
   // returns source of cvt
   auto getPrefetchSrc = [](Value v) -> Value {