Transpose Elementwise always if all branches can remove transpose

```mlir
%21 = stablehlo.convolution(%20, %arg4) dim_numbers = [0, 1, f, b]x[o, i, 1, 0]->[0, 1, f, b], window = {stride = [1, 1], pad = [[0, 0], [0, 0]], rhs_dilate = [1, 1]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64, precision_config = [#stablehlo, #stablehlo]} : (tensor<64x64x128x256xf32>, tensor<64x128x1x1xf32>) -> tensor<64x64x64x256xf32>
%22 = stablehlo.broadcast_in_dim %arg5, dims = [2] : (tensor<64xf32>) -> tensor<64x64x64x256xf32>
%23 = stablehlo.add %21, %22 : tensor<64x64x64x256xf32>
%24 = stablehlo.tanh %23 : tensor<64x64x64x256xf32>
%25 = stablehlo.transpose %24, dims = [3, 2, 1, 0] : (tensor<64x64x64x256xf32>) -> tensor<256x64x64x64xf32>

%51 = stablehlo.broadcast_in_dim %arg12, dims = [2, 1, 0] : (tensor<64x1x1xf32>) -> tensor<64x64x64x256xf32>
%52 = stablehlo.multiply %23, %51 : tensor<64x64x64x256xf32>
```

Start from the StableHLO convolution, elementwise operations, and transpose shown in the issue, and trace the compiler pass that decides whether to transpose elementwise branches. Check how the branches are analyzed for removable transposes. Done means the intended transpose behavior is applied when all branches can remove their transpose, with coverage for this operation pattern.

Written by the indexing model from the issue text.