Glow: Speed up our Travis tests

[bertmaher]

The ASAN configuration is close to 40 minutes, which is a bit crazy considering I can run the whole suite on my laptop in under four minutes. There’s a bit of obvious low hanging fruit:

• split test_unopt into its own Travis job
• make the test_unopt tests into actual cmake tests (rather than a custom command containing a big bash string) so that Ctest parallel level actually yields a speedup

[bertmaher]

Also is there a packaged protobuf we can install rather than building it from source? That adds a couple minutes, I think.

[rdzhabarov]

We've removed test_unopt from the ASAN run, now it takes ~23mins vs 40mins before (this fixes the first point :D)

[rdzhabarov]

Looking at tests themselves:

Test #16: BackendCorrectnessTest ...........   Passed  380.74 sec

out of 600s total.

Not related, but BackendCorrectnessTest is something we discussed with @jfix71 today earlier.

Jordan suggested that those tests are pretty much duplicates of operator tests and it makes sense to clean/delete or move some of them to operatorTest where appropriate.

[nadavrot]

@rdzhabarov @jfix71 The backend correctness test is the only place where we have a bunch of operators tested together. This gives the scheduler, the optimizer and the memory allocator something to work with. The operator tests usually just test one operator at a time.

[jfix71]

@nadavrot There are many tests in there that only test a single operator, though -- they just create placeholders, create operator, and save. So I was going to remove all of those, verifying that there is a test covering the operator in OperatorTest.

[rdzhabarov]

Some more speedups after caching building protobuf building. https://github.com/pytorch/glow/pull/1828

Hopefully the whole situation is going to be fixed after migration to circle ci.

[bertmaher]

Closing this in favor of the CircleCI task: #1836