Reimplement standalone score model.

[InfluenceFunctional]

We had previously had a standalone discriminator model which could be loaded and called by external libraries for crystal building and scoring.

• use standard pre-loading code
• load dataset, data standardization, model checkpoint
• currently sample analysis is done inside the standalone model - this is not ideal, but it is clean and self-contained
• simple tests