🐛 Bug

This issue is a collection of bugs where batched grad computation doesn't work. Most of these were discovered while turning batched grad testing on. Some of these manifest as crashes (😞 ), some as silently incorrect results, and some as loud error messages.

• test_linalg.py: test_cholesky_solve_autograd: The randomness in batched grad testing exposes unrelated silent correctness issues
• test_nn.py: test_broadcast_double_backwards_gpu: segfault
• test_nn.py: test_TransformerEncoderLayer_relu_activation: loud error message
• test_nn.py: test_CrossMapLRN2d: loud error message
• batched grad of torch.mm with a tensor of size (0, 0), as per @IvanKobzarev's comment

cc @ezyang @gchanan @zou3519 @bdhirsh @jbschlosser @albanD @gqchen @pearu @nikitaved @soulitzer