10-fold cross-validation works by taking the training set of labeled data and dividing it into 10 equal size subsets. 9 of the subsets are combined into the new training set and the remaining 1 subset is used for validation/testing, i.e. the model is trained on 90% of the original training set and tested on 10%.
This is performed 10 times (the folds) iterating over each of the 10 subsets so each subset is used for testing. A performance measure of the testing is performed on each iteration and after all iterations are completed, the average is calculated.
There is nothing called a "training fold" or "testing fold", a fold is an iteration of the process. There is also no subsets held aside during the process, all subsets are used in each iteration.
To create the learning curve you are talking about you could simply vary the size of the original training set and let the 10-fold cross-validation process run as it is. The number of records in the original training set is your measure of the training set size and the performance is the given average on completion of the cross-validation.