We should decorrelate XGBoost features after loading and prior to modeling

[mmcdermott]

We should support the options of:

• code/count -> event/count (once across all codes) & code/proportion_of_events (instead of the count of code occurrences, count of event occurrences and the proportion of event occurrences that are codes).
• value/count -> value/proportion_of_code (instead of the count of code occurrences with value, which is correlated with the code count, offer the proportion of code occurrences that have a value)
• value/sum -> value/mean (instead of the sum of the values, the mean of the values)
• value/sum_sqd -> value/variance (instead of the sum of the squares of the values, the variance of the values)
• Orthogonally, we should support decorrelating across window sizes, such that instead of having window sizes of [-7d, 0d] and [-1d, 0d], which are both mechanistically correlated as the second, shorter window is included in the first, larger window, we could construct windows of [-7d, -1d), [-1d, 0d] which are not mechanistically correlated.