Write tests for .log_abs_det_jacobian() methods

[fritzo]

I think we could do this by numerically approximating the Jacobian via finite difference. Alternatively we could stack a bunch of grad()s, computing the determinant as in this suggestion:

def test_jacobian(self):
    t = MyTransform()
    for x in my_test_points:
        y  = t(x)
        jacobian = torch.stack([grad(yi, [x], create_graph=True)[0] for yi in y])
        det_jacobian = torch.potrf(jacobian).diag().prod()**2
        log_abs_det_jacobian = torch.abs(det_jacobian).log()
        self.assertEqual(t.log_abs_det_jacobian(x, y), log_abs_det_jacobian)

It would be nice to make this test generic like the other tests in TestTransforms.

Note that this might not work for StickBreakingTransform because the det computation is over a non-square matrix since the source and target spaces have different dimensions.

[ssnl]

BTW, we now have torch.slogdet :)