answer 1 is non-sense and answer 2 is a much weaker/poorer version of answer 3; pretty much the same difference as in 1xN versus KxM cross-validation (with N=KxM).
Note (FYI, out of course scope): this later (asnwer 3) relates to boostrap methods in statistics.
Note (FYI, out of course scope): this later (asnwer 3) relates to boostrap methods in statistics.