rng = np.random.default_rng(seed=42)
    input_data = ops.convert_to_tensor(
        rng.uniform(low=0.0, high=1.0, size=static_input_shape).astype(
            np.float32
        )
    )

    def data_gen() -> repr_dataset.RepresentativeDataset:
      for _ in range(100):
        yield {
            'input_tensor': rng.uniform(
                low=0.0, high=1.0, size=static_input_shape
            ).astype(np.float32)

^bb0(%arg0: tensor<1x2xf32>):
  %cst_0 = arith.constant dense<[1, 0]> : tensor<2xi32>
  %0 = "tfl.quantize"(%arg0){qtype = tensor<1x2x!quant.uniform<u8:f32, 1.0>>}: (tensor<1x2xf32>) -> (tensor<1x2x!quant.uniform<u8:f32, 1.0>>)
  %1 = "tfl.dequantize"(%0): (tensor<1x2x!quant.uniform<u8:f32, 1.0>>) -> (tensor<1x2xf32>)
  %2 = "tf.Transpose"(%1, %cst_0): (tensor<1x2xf32>, tensor<2xi32>) -> tensor<2x1xf32>
  func.return %2 : tensor<2x1xf32>

At the same time, this behavior can be confusing for users as they can experience a failing test in a successful build.
To make the two APIs more uniform, we made `TestLauncher` also fail the build, which is a potential breaking change.
To continue the test execution even if a test task failed, Tooling API clients should explicitly pass `--continue` to the build.