Add an example to generate data on the circle.

[hongkai-dai]

Currently I found that I need a lot of training data to train a good score estimator.

Here is the noise conditioned score estimator with different size of training dataset

Here is the score estimator with a fixed noise level (sigma=0.05) with different size of training dataset

---

This change is