Fix MultiEpochSampler __len__() > 0 error

[JonFreer]

Fix for Issue #71

Caused when len(self.data_source) * self.num_epochs < self.start_itr * self.batch_size.

This happens once an entire pass of the dataset has occurred and then you resume training.

Fix: sets start_itr to the relative offset, resets to 0 after run through the dataset.

[szulm]

thanks