[HANDS-ON mistake] Unit 4 Policy gradient - mistake in exercise explanation

On the hands on section of unit 4 (policy gradient), when building the REINFORCE policy, the exercise has an intentional bug causing the error message
`ValueError: The value argument to log_prob must be a Tensor`.
The explanation says that this is error occurs because we should sample the probability distribution rather than argmax it - but that's not the reason this error occurs.
The error occurs because we use `np.argmax(Categorical(probs))` rather than `torch.argmax(probs)`.
Indeed, we would nevertheless like to substitute the argmax with sampling, but for different reasons, so the reasoning in the notebook is somewhat misleading.