example:

    ```
    TEST_F(ArrayGradTest, IdentityGrad) {
      TensorShape shape({5, 2});
      auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(shape));
      auto y = Identity(scope_, x);
      RunTest(x, shape, y, shape);
    }
    ```

NOTE: There are some ops that require features from the C++ API that are not yet
implemented.

* `--tags`: Comma-separated list of tags for loading SavedModel. Ignored for
  MLIR input.
* `--input-arg-shapes`: A string representation of input argument shapes for
  'main' entry-point, separating tensors with ':', dimension with ',', and
  using '?' for unknown sizes. For example, `input-arg-shapes=1,2::1,?`
  expresses argument shapes `[1,2]`, `[]` and `[1,?]`.
* `--e`: Elide large elements attrs while dumping the output StableHLO.

    new op is presented as a single node in the graph, thus optimization passes
    and kernels can easily be specialized to this op for better performance.
*   *Automatic shape/type inference support*: No shape functions are required
    for the new op;
*   *Automatic gradient support (WIP)*: The user doesn't need to author gradient
    a function of the op for training.

### Use Cases