[wasm] Add ResizeBilinearGrad and ResizeNearestNeighborGrad kernels

[chunnienc]

@mattsoulanille and reviewers: This compiler explorer example https://godbolt.org/z/zGr56TWeT shows that the shape (helper) class and lambda argument are zero-cost in our usage as we discussed. You can see that the generated assembly codes for Operator1 (with these features) and Operator2 (native implementation) are the same.

To see the logs from the Cloud Build CI, please join either our discussion or announcement mailing list.

---

This change is 

[mattsoulanille]

LGTM. The shape abstraction makes indexing much easier to read!

[pyu10055]

Thank you, nice Shape class.

Reviewable status: complete! 2 of 1 approvals obtained (waiting on @chunnienc)

---

tfjs-backend-wasm/src/cc/shape.h line 57 at r1 (raw file):

  }

  inline T& operator[](size_t i) { return data_[i]; }

boundary check on i?

[chunnienc]

Thank you, nice Shape class.

Reviewable status: complete! 2 of 1 approvals obtained (waiting on @chunnienc)

tfjs-backend-wasm/src/cc/shape.h line 57 at r1 (raw file):

  }

  inline T& operator[](size_t i) { return data_[i]; }

boundary check on i?

The operator[] does not do boundary checking, while the .at(i) method checks the boundary and can throw an std::out_of_range error. This is a convention in C++ STL and almost all containers (std::vector, std::array, etc) implement this behavior. That's also why I expose two methods for different use cases.
Ref:
https://en.cppreference.com/w/cpp/container/array/operator_at
https://en.cppreference.com/w/cpp/container/array/at