CPU/GPU consistency for pow

[jgartman]

Description

This PR makes the output of the pow function more consistent on WebGL vs. the CPU tensorflow/tfjs#350. This change will cause the function to output nan when a negative base is taken to a fractional exponent.

---

For repository owners only:

Please remember to apply all applicable tags to your pull request.
Tags: FEATURE, BREAKING, BUG, PERF, DEV, DOC, SECURITY

For more info see: https://github.com/tensorflow/tfjs/blob/master/DEVELOPMENT.md

---

This change is 

[nsthorat]

Thanks for the PR! Just curious, do you know if this is how TensorFlow python behaves?

---

Review status: 0 of 1 LGTMs obtained

---

Comments from Reviewable

[jgartman]

TensorFlow python does return nan for a negative base taken to a fractional exponent however there are still some other differences with the tfjs implementation. TF is stricter about both the arguments having the same type but based on this discussion #1129 it doesn't seem like that's a concern. TF will also throw an error when an int32 base and exponent are used and the exponent is negative but again I thought that was stricter than necessary here. Just to give a little background, the primary motivation for this PR was this discussion tensorflow/tfjs#346 about how to fix the gradient for pow.

Note that the testing I did with TF python was CPU only.

---

Review status: 0 of 1 LGTMs obtained

---

Comments from Reviewable

[nsthorat]

---

Review status: complete! 1 of 1 LGTMs obtained

---

Comments from Reviewable

[nsthorat]

Thanks @jgartman!! P.S. you should still shoot me an email, I can't find your email anywhere :)

[jgartman]

Will do @nsthorat and if you ever need my email its my firstname.lastname at gmail.