Fix calculated variables being left out of SingleYearDataset objects

[juaristi22]

Fix #22

[baogorek]

My only comments related to documentation (e.g., docstrings) and are non-blocking. I did have a bit of trouble figuring out exactly what was going on. Finally I get it. This Claude Code summary was really nice I thought:

● The PR fixes a bug where pre-computed income values were being lost during dataset minimization.

  When creating smaller dataset subsets for calibration, the code was only copying "input variables" (raw data like age, state) but dropping "calculated variables" (pre-computed
  values like employment_income, self_employment_income stored in the original dataset).

  This resulted in minimized datasets with all zeros for income fields, making them useless for calibration.

  Maria's fix:
  1. Identifies which variables in the dataset are calculated (not inputs)
  2. Explicitly preserves these calculated variables and their values when creating subsets
  3. Ensures income data remains intact throughout the calibration pipeline

  The tests confirm that employment_income, self_employment_income, and weekly_hours_worked now retain their non-zero values after minimization.

If there's any risk that other users are going to hit this in the future and not know what happened, is there any way to warn them proactively?