Bug in `OperationEquivalence` (breaks `-cse` on `linalg.index`)

MLIR
19

I found another CSE problem with a different operation:

func.func @masked_reduce_minf_f32(%arg0: vector<16xf32>, %mask : vector<16xi1>) -> (f32,f32) {
  %2 = vector.reduction <minnumf>, %arg0 : vector<16xf32> into f32
  %0 = vector.mask %mask { vector.reduction <minnumf>, %arg0 : vector<16xf32> into f32 } : vector<16xi1> -> f32
  return %0, %2 : f32, f32
}

If you run CSE, you get this:

  func.func @masked_reduce_minf_f32(%arg0: vector<16xf32>, %arg1: vector<16xi1>) -> (f32, f32) {
    %0 = vector.reduction <minnumf>, %arg0 : vector<16xf32> into f32
    %1 = vector.mask %arg1 { vector.yield %0 : f32 } : vector<16xi1> -> f32
    return %1, %0 : f32, f32
  }