Purging handles the training rows before a test fold. The embargo handles the rows just after it. Because returns and features are autocorrelated, a training row dated one day after the test fold still carries information about the end of that fold.
The convention is to embargo about 1% of the total sample length, or the length of the feature lookback, whichever is larger. On 5,000 daily observations with a 60-day feature window, embargo 60 days after each test fold.
It feels wasteful to throw away data, and it is: with 5 folds and a 60-day embargo you lose 300 observations. That cost buys you an accuracy estimate you can believe, which is worth far more than 6% more rows.
Related: purged-cross-validation, cross-validation, data-leakage