Add NonMaxSuppressionV5 to tfjs-node

[lina128]

• Add the op using modular pattern
• Create new folder for all new kernels

Note: Tested with latest core at e5a0c84

---

This change is 

[nsthorat]

Awesome! Can you update the dependency to 1.5.0 to have the tests run?

Reviewable status: 0 of 1 approvals obtained (waiting on @dsmilkov, @kangyizhang, @lina128, @nsthorat, and @pyu10055)

---

tfjs-node/src/kernels/non_max_suppression_v5.ts, line 41 at r1 (raw file):

    const {maxOutputSize, iouThreshold, scoreThreshold, softNmsSigma} =
        attrs as NonMaxSuppressionWithScoreAttrs;
    const maxOutputSizeTensor = scalar(maxOutputSize, 'int32');

to do this modularization properly we shouldn't have a dependency on tensors here, we should just operate entirely dataid's. This probably means some refactoring of executeMultipleOps

[dsmilkov]

LGTM!

Reviewed 3 of 3 files at r1.
Reviewable status: complete! 1 of 1 approvals obtained (waiting on @dsmilkov, @kangyizhang, @nsthorat, and @pyu10055)

---

tfjs-node/src/kernels/non_max_suppression_v5.ts, line 41 at r1 (raw file):

Previously, nsthorat (Nikhil Thorat) wrote…

to do this modularization properly we shouldn't have a dependency on tensors here, we should just operate entirely dataid's. This probably means some refactoring of executeMultipleOps

You or I can do that in a separate PR, since the node backend needs refactor (we never refactored it to disentangle tensors from tensor handles)