Test error weighting by time

[dfsnow]

Per @Douglasmsw, let's test weighting the model errors by time, such that the errors of more recent sales are larger. This should be accomplishable with the weights parameter built into LightGBM (though it might be a bit faffy). There's also a weights argument built into Lightsnip: https://github.com/ccao-data/lightsnip/blob/master/R/lightgbm.R#L135

As a first pass, let's try simply weighting all post-2020 sales with a fixed constant. If that works (i.e. the predicted values change) we can then:

• Test out different weighting functions (e.g. Gaussian, logistic, etc.)
• Figure out a way to search the parameter space of the aforementioned functions to find the most appropriate weights
• See also Test propensity weights as case weights for ultra high value properties #52