Add a dataset module

[nudles]

Data loading is an important part of DL training, which could be slow and become a bottleneck if not implemented well. The tasks include

1. implement dataset classes for common benchmark datasets to make them easy to access within SINGA (e.g., without manual downloading).
2. implement common preprocessing operations
3. implement parallel data loading for higher efficiency

[nudles]

Code from the data module may be reused. https://github.com/apache/singa/blob/master/python/singa/data.py

[nudles]

And https://github.com/apache/singa/blob/master/python/singa/image_tool.py