Currently, multi_tensor_apply causes an illegal memory access due to
an overflow in the `size` field of `TensorListMetadata`. This can be
reproduced using the following standalone script:

```python
import torch, amp_C
from apex.multi_tensor_apply import multi_tensor_applier
multi_tensor_adam = amp_C.multi_tensor_adam

size = 2**32+1
g_32 = [torch.zeros(size, dtype=torch.float32, device='cuda')]
p_32 = [torch.zeros(size, dtype=torch.float32, device='cuda')]
m_32 = [torch.zeros(size, dtype=torch.float32, device='cuda')]
v_32 = [torch.zeros(size, dtype=torch.float32, device='cuda')]
_dummy_overflow_buf = torch.zeros(1, dtype=torch.int32, device='cuda')

multi_tensor_applier(multi_tensor_adam, _dummy_overflow_buf, [g_32, p_32, m_32, v_32], 0.0, 0.9, 0.95, 1e-08, 1, 1, 1, 0.1)
print(g_32)
```