Avoid Segfault for scalar shapes.

Calling tensor::FromElementsOp with an empty vector of elements and no type
causes a segfault. We need to let the FromElementsOp know which scalar type it
should have.
Also add back the DynamicBroadcastInDimOp canonicalization patterns, which
previously prevented this bug from happening.
Add a regression test that demonstrates the bug.

PiperOrigin-RevId: 417561444
Change-Id: I6d1d6cfb71aabbad6102422625a00bbe253ac95a

tensorflow/compiler/mlir/tfrt/jit/transforms/tf_cpurt_symbolic_shape_optimization.cc

@@ -157,6 +157,10 @@ llvm::Optional<Value> simplifyBroadcast(ShapeComponentAnalysis& analysis,
     shapes_found.push_back(*found_shape);
     maxRank = std::max(maxRank, found_shape->size());
   }
+  if (maxRank == 0) {
+    return Value(builder->create<tensor::FromElementsOp>(
+        loc, shapes[0].getType(), SmallVector<Value>()));
+  }
 
   SmallVector<const ShapeComponentAnalysis::SymbolicExpr*> joined_dimensions(
       maxRank);

tensorflow/compiler/mlir/tfrt/python_tests/regression_tests/scalar_broadcast.mlir

@@ -0,0 +1,11 @@
+builtin.func @test(%V__0 : tensor<i1> { python_test_attrs.static_type = tensor<i1> }, %V__1 : tensor<f32> { python_test_attrs.static_type = tensor<f32> }, %V__2 : tensor<f32> { python_test_attrs.static_type = tensor<f32> }) -> tensor<f32> {
+  %0 = "tf.Cast"(%V__0) : (tensor<i1>) -> tensor<i1>
+  %1 = "tf.Selu"(%V__2) : (tensor<f32>) -> tensor<f32>
+  %2 = "tf.NextAfter"(%1, %V__2) : (tensor<f32>, tensor<f32>) -> tensor<f32>
+  %3 = "tf.Elu"(%2) : (tensor<f32>) -> tensor<f32>
+  %4 = "tf.Cosh"(%3) : (tensor<f32>) -> tensor<f32>
+  %5 = "tf.Elu"(%4) : (tensor<f32>) -> tensor<f32>
+  %6 = "tf.Div"(%V__1, %5) : (tensor<f32>, tensor<f32>) -> tensor<f32>
+  %7 = "tf.Select"(%0, %6, %V__1) : (tensor<i1>, tensor<f32>, tensor<f32>) -> tensor<f32>
+  return %7 : tensor<f32>
+}