rinet 0.2.0

Backend: torch replaces Python/TensorFlow

The package no longer requires Python, TensorFlow, Keras, or scikit-learn. Models now run through the torch R package. After installing rinet, run torch::install_torch() once to download the backend.
This is an infrastructure change only: predictions from the v2.0 model are identical to rinet 0.1.x up to floating-point precision.

The default 1D model is now v2.1. It has the same inputs and outputs as v2.0 but was retrained on revised simulated data: pathological component shifts and widths are drawn relative to the reference component (instead of in absolute units with clamping), and the training set is five times larger. This gives more even coverage of mixtures in which the pathological component is far from, or narrower than, the reference.
The model described in the published paper is still available as model_version = "v2.0" in predict_rinet() and predict_rinet_1d().
Each prediction now includes a model_version element.
The 2D model is unchanged (v2.0).
The updated simulation and training code is available at https://github.com/jackgle/RINet.

RIbench benchmark (Ammer et al., 2022; 5,760 simulated test sets, 10 analytes; absolute z-score deviation, lower is better): overall score improved from 0.395 to 0.356.
- Largest gains at high pathological fractions, e.g. normally distributed analytes improved from 0.523 to 0.379.
- Some loss at low pathological fractions for skewed analytes: skewed 0.364 to 0.371, heavily skewed 0.249 to 0.267.
Liver function test dataset (10 analytes x 2 sexes, normalized error against the direct method on reference-labelled patients): mean error 0.061 (v2.0) vs 0.067 (v2.1), within the variation expected for 20 cases.