First, it is useful to realize that your question is about statistical learning and not quantum computing. Your question better fits the Cross Validated stack.
But since your code contains a tiny bit of quantum computation along with a classical neural network I give a brief answer.
This plot doesn't say anything except that the current random split of data into training and validation sets produces a validation error slightly higher than the training error. Nothing wrong with it.
Purely by chance, the overall dataset may be split so that a bad model may produce a good validation score or vice versa.
This means that a single split of data really doesn't say much about the performance of a model. What actually needs to be done is to gather various training and validation statistics over multiple data splits. This is usually achieved by the means of well-known methods such as cross-validation or even bootstrapping. Only once you consider statistics given by the aforementioned procedures you will be able to tell if your final choice model overfits.