ci: pin to non-broken nightlies

[wchargin]

Summary:
Due to tensorflow/tensorflow#34427.

This unblocks our CI until the upstream problem is identified and fixed.

Test Plan:
Fingers crossed.

wchargin-branch: pin-20191118-nightly

[stephanwlee]

Read your issue on TensorFlow. Thanks for the thorough investigation!

[wchargin]

Well, okay, now I’m stumped. I installed the exact same pip freeze env
(with tf-nightly==2.1.0.dev20191118) into a new virtualenv on a new
Ubuntu 14.04 machine, and cannot reproduce the problem there. But Travis
is still broken. I’m going to clear the caches on Travis and re-trigger
the two problematic builds. If that doesn’t work, I’ll try pinning to
2.1.0.dev20191024 instead, as in #2834.

[wchargin]

Cleared cache for #2947. There doesn’t seem to be a cache for master
for some reason. Re-running now.

[wchargin]

Did not suffice. I will clear all repository-wide caches then try
2019-11-18 nightlies again. (Rolling back forward because some debugger
tests are failing in ways that may or may not be related to recent
TensorFlow changes, and I want to rule out that possibility.)

[wchargin]

Okay, after deleting the caches for the release branches Travis deigned
to show me that there was indeed a cache for master, so I have now
actually deleted that. Retrying presently.

[wchargin]

Still SIGABRTing in smoke test…

*** Error in `python': free(): invalid pointer: 0x000000000255f3c0 ***
Aborted (core dumped)

[wchargin]

Closing and reopening per Travis support suggestion…

[wchargin]

Closing and reopening again per Travis support suggestion (they’ve made
changes on the backend)…