Add random seed replicates to predict step

[jds485]

Noted in #125, here:

Random draws of samples into the training and testing sets have resulted in different outcomes for training using all available gages vs. using gages only within the region for which testing will be completed. Adding several replicates and comparing the resulting distribution of RMSEs is one way around this problem.

I think the replicates should use the same selected attributes (from correlation and Boruta screening) and same hyperparameters (to reduce computation times). So, the only difference between replicates would be the dataset used.

With replicate models, we should edit:

• feature importance plots to display the average value over all replicates (can add error bars)
• Model RMSE comparison barplots (e.g., p6_compare_RMSE_RF_png) should show average test error and average validation error over all replicates (can add error bars)
• predicted vs. observed scatterplots to show distributions over all replicates (not sure how is best to show this - would be too crowded to add error bars).
• residual maps based on average residual over replicates