if by "Guard interval" you mean the "Zero Subcarriers" in OFDM symbols , i should notice that the zero padding makes the filtering requirements easier to be satisfied in digital domain(if we don't zero pad the symbols , we would need a LPF with a very sharp transition band, which is difficult to implement digitally).
as you said in your post , another way for channel seperation is to use oversampling.
Yes, typical the channel filtering requires minimum oversampling of 2x of the symbol sampling frequency to get more relaxed requirements for alias and far away blocking filtering. But exact this reason also state that guard carriers are not sufficient for channel filtering alone. So there is NO specific reason to use guards if oversampled filtering is required anyway.
It remains: What is the true reason for guard carriers?
After the IFFT part, windowing(filtering) is done. But the filter is not having a sharp cutt-off frequency. It will have some roll-off factor if the frequencies used in the OFDMA symbol falls in this region, then it will be attenuated. There fore they leave guard band in left as well as right side.
Refer to the attachment if you have doubt on where windowing is done.
So for OFDM transmission the signal have to be send to the DAC's. This DAC's operate at 2-8x symbol sampling frequency. Then there is analog aliasing. The interpolation filtering reduce the images in between fs/2 and fdac/2.
What is the net benefit of a little, small guard band of subcarriers, for the effort of interpolation and analog filtering in the transmitt path?
Is the 2^k of used OFDM carriers a better approach for OFDMA than having a little trade gain because of the guards?