Lower learning rate for TF cifar integ test (#265)

We've been seeing a lot of errors that say failure due to NaN loss.
One suggestion from https://stackoverflow.com/questions/40050397/deep-learning-nan-loss-reasons
is to use a lower number for learning rate.

Author: laurenyu

tests/data/cifar_10/source/resnet_cifar_10.py

@@ -33,8 +33,8 @@
 BATCH_SIZE = 1
 
 # Scale the learning rate linearly with the batch size. When the batch size is
-# 128, the learning rate should be 0.1.
-_INITIAL_LEARNING_RATE = 0.1 * BATCH_SIZE / 128
+# 128, the learning rate should be 0.05.
+_INITIAL_LEARNING_RATE = 0.05 * BATCH_SIZE / 128
 _MOMENTUM = 0.9
 
 # We use a weight decay of 0.0002, which performs better than the 0.0001 that