Canonicalize flow.tensor.clone to keep it more local to usage

[benvanik]

Now that we are performing in-place operations there are flow.tensor.clones appearing to preserve correctness. In cases of wide fan-out these cause some pathological behavior:

      %22 = flow.dispatch @serving_default_dispatch_21::@serving_default_dispatch_21[%c384, %c4, %c1]() : () -> tensor<1x4x384xf32>
      %23 = flow.tensor.clone %22 : tensor<1x4x384xf32>
      %24 = flow.tensor.clone %23 : tensor<1x4x384xf32>
      %25 = flow.tensor.clone %24 : tensor<1x4x384xf32>
      %26 = flow.tensor.clone %25 : tensor<1x4x384xf32>
      %27 = flow.tensor.clone %26 : tensor<1x4x384xf32>
      %28 = flow.tensor.clone %27 : tensor<1x4x384xf32>
      %29 = flow.tensor.clone %28 : tensor<1x4x384xf32>
      %30 = flow.tensor.clone %29 : tensor<1x4x384xf32>
      %31 = flow.tensor.clone %30 : tensor<1x4x384xf32>
      %32 = flow.tensor.clone %31 : tensor<1x4x384xf32>
      %33 = flow.tensor.clone %32 : tensor<1x4x384xf32>
      %34 = flow.tensor.clone %33 : tensor<1x4x384xf32>
      %35 = flow.tensor.clone %34 : tensor<1x4x384xf32>
      %36 = flow.tensor.clone %35 : tensor<1x4x384xf32>
      %37 = flow.tensor.clone %36 : tensor<1x4x384xf32>
      %38 = flow.tensor.clone %37 : tensor<1x4x384xf32>
      %39 = flow.tensor.clone %38 : tensor<1x4x384xf32>
      %40 = flow.tensor.clone %39 : tensor<1x4x384xf32>
      %41 = flow.tensor.clone %40 : tensor<1x4x384xf32>
      %42 = flow.tensor.clone %41 : tensor<1x4x384xf32>
      %43 = flow.tensor.clone %42 : tensor<1x4x384xf32>
      %44 = flow.tensor.clone %43 : tensor<1x4x384xf32>
      %45 = flow.tensor.clone %44 : tensor<1x4x384xf32>
      %46 = flow.dispatch @serving_default_dispatch_22::@serving_default_dispatch_22[%c384, %c4, %c1](%21, %22) : (tensor<4x384x384xf32>, tensor<1x4x384xf32>) -> %22
      // uses of %23-%46 throughout the model

Some simple code motion will help here: sinking the clones to immediately prior to their use will shorten the lifetime of these transient tensors.