Modularize remaining unary ops

[tafsiri]

modularize round, rsqrt, sigmoid, isNan, isInf, isFinite, softplus, sqrt, step

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

---

This change is 

[lina128]

Thank you Yannick!

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @annxingyuan, @lina128, and @tafsiri)

---

tfjs-core/src/gradients/IsInf_grad.ts, line 27 at r1 (raw file):

  kernelName: IsInf,
  gradFunc: (dy: Tensor) => {
    // TODO(nsthorat): Let gradients be null for cases where we want to stop

Curious, do we still want to do this?

[tafsiri]

@lina128 good question. I don't really know, cc @nsthorat about the TODOs in these gradients

[nsthorat]

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @annxingyuan and @lina128)

---

tfjs-core/src/gradients/IsInf_grad.ts, line 27 at r1 (raw file):

Previously, lina128 (Na Li) wrote…

Curious, do we still want to do this?

yes I think we should do this, its a perf optimization. Right now we pass zeros backwards and do more work than we need to (while null can signal stop propagating the gradient backwards altogether)