ci: pin to non-broken nightlies

[wchargin]

Summary:
Travis builds are failing on today’s tf-nightly and not on the other
environments (1.15.0rc3 and notf). The error is in the smoke test:

+python -c
import tensorboard as tb
tb.summary.v1.scalar_pb('test', 42)
tb.summary.scalar('test v2', 1337)
from tensorboard.plugins.beholder import Beholder, BeholderHook
from tensorboard.plugins.mesh import summary
2019-10-24 19:50:45.339101: W tensorflow/stream_executor/platform/default/dso_loader.cc:55] Could not load dynamic library 'libcuda.so.1'; dlerror: libcuda.so.1: cannot open shared object file: No such file or directory
2019-10-24 19:50:45.339143: E tensorflow/stream_executor/cuda/cuda_driver.cc:351] failed call to cuInit: UNKNOWN ERROR (303)
*** Error in `python': free(): invalid pointer: 0x00000000013e6340 ***
Aborted (core dumped)
+cleanup
+[ -n set ]
+rm -rf /tmp/tmp.E6thqwE4O7

I can’t reproduce this locally on a virtualenv with these packages:

pip: tb-nightly==2.1.0a20191024
pip: tf-nightly==2.1.0.dev20191024
python: tb.__version__ == 2.1.0a20191024
python: tf.__git_version__ == v1.12.1-16634-g0e90b26
python: tf.__version__ == 2.1.0-dev20191024

Downgrading nightly in the hope that this will unblock PRs, and will try
to understand the underlying issue later.

Test Plan:
See what Travis thinks.

wchargin-branch: pin-20191023-nightly