Previously, running `KPM predict` on a single reactant/product pair would cause KPM to fail due to NumPy loading in the enthalpy as a 0-dimensional array which couldn't be indexed. Fixed by enforcing a 1D array be loaded.

KPM could not run headless (without a `DISPLAY` environment variable set) due to it loading the `TkAgg` backend for Matplotlib. This has been fixed, so if `DISPLAY` is not set or empty, KPM will instead load the `Agg` backend, allowing for headless running of both training and rediction modes.

Calling `KPM predict` with a direction of 'forward' or 'backward' was failing due to an incorrectly set `num_reacs` in `calc_diffs()`. This has now been set correctly based on the requested prediction directionality.

…odel into main

Fixed ensemble training procedure to enable uncertainty quantification on Ea predictions by taking the variance over `nn_ensemble_size`. Added support for outputting uncertainty in both train/test correlation plots and prediction outputs.

Also added ability to choose whether Eacts should be internally normalised.