Add information about non-differentiable points to grad tests

[soumith]

FAIL: test_ReLU (__main__.TestNN)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "test_nn.py", line 287, in <lambda>
    setattr(TestNN, test_name, lambda self,test=test: test(self))
  File "/data/users/soumith/pytorch/test/common_nn.py", line 533, in __call__
    self._do_test(test_case, module, input)
  File "test_nn.py", line 33, in _do_test
    test_case.check_jacobian(module, input, self.jacobian_input)
  File "/data/users/soumith/pytorch/test/common_nn.py", line 433, in check_jacobian
    PRECISION
AssertionError: 0.20518362972049153 not less than or equal to 1e-05

ERROR: test_maskedSelect (__main__.TestTorch)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "test_torch.py", line 1887, in test_maskedSelect
    self.assertEqual(dst, torch.Tensor(dst2), 0)
  File "/data/users/soumith/pytorch/test/common.py", line 66, in assertEqual
    max_err = max(max_err, abs(x[index] - y[index]))
TypeError: bad operand type for abs(): 'DoubleTensor'

[apaszke]

maskedSelect test is fixed in a8e816f. As for the ReLU test, it appears to happen around once in 15000 tests and it's due to the non-differentiability of ReLU at 0 (it happens when randn returns a number in interval of [-1e-6, 1e-6]). If we want to make the tests failproof we would need to mark the non-differentiable points of each function/module and make sure no input is too close to them.