When we augment data for training are we also changing the distribution of data and if its a different distribution why do we use it to train a model for original distribution ?
Yes you do change the distribution of your training data if you modify it (for example by augmenting it with rotated versions of images that were in an original training set of images).
This is fine because, typically, our goal of training is not to get a model with high performance on the dataset we happened to collect as training data (e.g. a bunch of natural images). Our goal is to train a model that generalizes well to new data outside of the training data's distribution.
Typically, the training data is only a sample of the distribution we're actually interested in. For example, we'll be interested in making accurate predictions for all natural images in the entire world. That is a distribution that would likely include rotated variants of all images in our training set. So, if we augment our training set by adding such rotated variants, we expect to modify our training data distribution in such a way that it actually gets a little bit closer to the distribution we're interested in (all natural images).