Add absolute tolerance to test_net.py to prevent random Travis fails

[Noiredd]

Follow up on issue #5960 - fix by adding absolute tolerance of 1e-5 to test_forward_start_end() and test_backward_start_end() in test_net.py. I iterated python -m unittest discover -s caffe/test -p "test_net.py" over the current master 500 times and no random fails occurred.

Additionally I removed semicolons that ended each line in those two functions - it looked weird.

[shaibagon]

LGTM

[Noiredd]

Why did you opt for adding atol instead of increasing slightly the random input numbers?

I explained that in the comment under the issue itself. This test does not only concern randomly generated data, but also weights in the InnerProduct layer of the test net. We can increase the data slightly but the weights (drawn from a normal distribution) can still roll such that some output will be small.

[shaibagon]

@Noiredd what holds you from merging this PR?

[shelhamer]

Thanks for addressing these spurious flaky tests!