[WASM] Add PadV2 kernel

[dsmilkov]

Implement Padding, needed by almost all models. I added a fast implementation for up to rank-4 tensors, and a generic slow implementation for any rank.

FEATURE

---

This change is 

[annxingyuan]

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @annxingyuan, @dsmilkov, and @tafsiri)

---

tfjs-backend-wasm/src/cc/kernels/PadV2.cc, line 137 at r1 (raw file):

  typed_memset<T>(out_data, pad_value, out_size);

  for (size_t i = 0; i < in_size; ++i) {

Would we be able to memcopy x_data in the generic case as well by looping over x_shape and paddings?

[dsmilkov]

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @annxingyuan and @tafsiri)

---

tfjs-backend-wasm/src/cc/kernels/PadV2.cc, line 137 at r1 (raw file):

Previously, annxingyuan (Ann Yuan) wrote…

Would we be able to memcopy x_data in the generic case as well by looping over x_shape and paddings?

Not sure how to do memcpy and work with arbitrary rank. Either way, this slow_pad_nd is a fallback and shouldn't be used by our models which have padding for rank<=4