Separate fitting and prediction data

[zhengp0]

Currently we store fitting and prediction data in one data frame. While this is convenient for our current use, it have several drawbacks

• We will have data set that has multiple observations per demographics combination. And this will potentially make the prediction step more complicated and expensive than necessary
• For weave model it requires all data frame doesn't have NaN, and we need to manually set the NaN in the residual where we do not have observation but want prediction to 1.