ENH: Add workflow to train model

[deshmukg]

• Closes Create training and prediction workflows #11.
• Added train.py that contains Standardizer to standardize the outputs, Model class to initialize a model and train/validate/predict.
• Added partition_structure_by_layer to partition structures based on layer number rather than z-coordinate. This works better than partition_structure since different alloys have different lattice constants and hence, different slab heights. Therefore, the latter leads to incorrect graphs if different alloys are present in the dataset. All the key classes and functions now use partition_structure_by_layer by default. The interlayer spacing is calculated by sorting the list of unique z-coordinates of a slab in ascending order and taking the difference of the first two elements. This is essentially the calculation of the interlayer spacing of the bottom two layers. This method works if the bottom few layers are constrained in a calculation.
• Converted lists to store layers in MultiGCN to nn.ModuleList objects. This is required to run the model on a GPU, since Python lists cannot be transferred to a GPU.
• Added Deal with __init__.py when making package #21.