Flaky tests when using `model_probably_good`

[thisac]

Tests using model_probably_good are flaky and sometimes fail.

..................................................F....
======================================================================
FAIL: test_SkipLinear_different_dim (test_nn.TestLinear.test_SkipLinear_different_dim)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/Users/distiller/project/tests/test_nn.py", line 86, in test_SkipLinear_different_dim
    self.assertTrue(model_probably_good(model, (din,), (dout, )))
AssertionError: False is not true

Should look into either switching them out for non-random tests or make sure that they pass with much more certainty.

[kevinchern]

How did you trigger this failure? The test is unlikely to have failed from the random test. The input consists of a 100x99 tensor of floats. The probability of failure is exponentially small. This test needs to be improved for at least two reasons: 1) this test isn't helpful in isolating the problem, e.g., is it due to having zeros in the output or is it likely bounded? 2) not reproducible.

[thisac]

IIRC this happened on occasion in CI. Assumed it to be flaky based on model_probably_good.