Created by: zhiqiu

Fix a bug when using varbase.__getitem__ in python2 with long int, e.g.:

The correct result should be [2L, 80L, 16000L].

Plus, original __getitem__ does not raise errors when called with float and other types, while numpy and torch does. e.g.: var[1.1] or var[:, 1.1:]

This PR enhance the error handling. e.g.: