Dynamic loading on data

[penfoldtj]

We can generate some large training sets now - but then run into memory issues. It would be good to implement an option for dynamic loading of batches during learning - so everything is not held in memory all the time.

• Covert multiple .xyz files into a single HDF5 file
• Implement a class to load structures dynamically from HDF5
• Integrate the new dataset with Dataloader for batch loading