The current clip_grad_norm_ and clip_grad_value_ takes Iterable[Tensor] as input.

Is it better to support clip_grad_* to accept a single tensor as the input, and do clip on that tensor? For now, if we feed a tensor, it will be treated as an Iterable, which will result in no grads to clip.

Or, instead, should we raise a warning or error when the input is a tensor?