Understand why epsilon checks are necessary in packed division and add test exposing root issue.

[annxingyuan]

This issue (#1321) prompted this fix (tensorflow/tfjs-core#1603) of checking whether the inputs to packed div are less than epsilon and setting the output to 0 if so. We need to understand why it's necessary to check for less than epsilon, rather than equals 0.

We also need to understand why the inputs were 0's. I thought it was because they were in out-of-bounds channels, but an alternative fix of checking whether rc channels were out of bounds did not fix the problem. So that makes it seem like the inputs were 0's in the data, but the unpacked version of division did not produce NaNs.