subsample vs colsample_bytree

A LightGBM model has 200 features, many of which are correlated groups (e.g., you have 10 slightly different versions of "user activity last N days" for N = 1, 3, 7, 14, 30, 60, 90, 120, 180, 365).

1. Explain the difference between subsample and colsample_bytree.
2. Why are correlated features a problem for gradient boosting even when each feature individually is predictive?
3. Which of the two parameters is more important for this specific dataset, and why?