Zero stuffing does not insert additional frequencies and the frequencies above the original signal frequency are not present in the original signal; however, because the original signal is a set of samples, there are convolved images in the frequency domain and those higher frequencies will be taken up by the first duplicate image in this case. It simply increases the amount of frequencies that are sampled, i.e. the period of sampling in the frequency domain widens and ends up including the next image of the original signal in the frequency domain, which is known as imaging.
Because the original signal has a finite period in the time domain, it must therefore have discrete frequency samples in the frequency domain (which is shown on row 2; the dotted line indicates an impulse envelope and the ellipses indicate that the period repeats infinitely). The signal has a total bandwidth of Fs or a Fmax of 1/2 Fs, therefore it needs to be sampled at at least Fs as the Nyquist rate, which is whatever the bandwidth happens to be. When this time domain signal is sampled on row 3 at every Ts seconds, it convolves the frequency domain every Fs.
On the 4th row, we upsample a time domain signal 2x which is already a set of samples and hence already has infinite images in the frequency domain. Because it is multiplying a set of impulses with a denser set of impulses, there will be 0s where the impulses don't intersect (the dotted line shows the envelope of the impulses and not a continuous signal, so it is actually 0 between the impulses) (in this case every other sample will be a 0 because we are upsampling 2x. Ts is now half the Ts of the original signal). The frequency domain of the signal will now be convolved every 2Fs. The new Fs is 2x the original. The time domain samples will just be the original samples but with 0s interspersed between the original samples.
The resulting frequency domain is identical, except Fs now covers a 2x larger window of the frequency domain. This means that you need a low pass filter at the frequency of the original Fs to remove the unwanted frequencies to get the resulting Fs window you would have got from sampling the original continuous signal at that sampling rate as opposed to a set of samples.
Sampling the original time domain signal 2x would have yielded:
You don't need a low pass filter here because the images are greater than the Nyquist frequency.
When you filter out the imaging of the upsample in the first scenario, the frequency domain resembles the above scenario. It is called interpolation, because it turns each 0 into an interpolation between the points either side, to match identically the time domain of above scenario. The zero stuffing is the prerequisite part of interpolation.
All in all, upsampling is the process of zero stuffing and interpolating (filtering) a set of samples of signal to give the set of samples a higher sampling rate, as if they had been taken from the original analogue signal at that higher sampling rate. The underlying frequency components in the signal does not change. You are just sampling it at a higher rate.
The DFT of the zero stuffed samples is the original frequency domain samples and another set of samples the same size added to the end which sample an image. If you make the samples of the image all 0 and then perform an IDFT, the 0s in the resulting set of time domain samples now become interpolated points.
The frequency domain is a series of infinite images which has been multiplied by the filter, which is why the rolloff matters, because the filter is not being multiplied by 0, but by images (and the side lobes of the frequency domain sinc impulses). It is a series of infinite images because the time domain was a series of impulses, which were windowed to the signal length (which merely causes a convolution around the impulses in the images) and then the multiplication with the filter removes all of these images and the most possible of the 2 immediately adjacent images either side.
The original signal and the upsampled signal have the same window size and therefore the impulses representing the samples in the frequency domain have the same shape and size. The impulses in the time domain of the upsampled signal are smaller and hence the filter in the frequency domain is wider to match the sampling frequency and so is the ZOH DAC frequency response. The resulting pulse shape filter is wider and this allows for some side lobes of the frequency domain impulse sincs to be included in the output because they are not removed by the filter (this is identical to the scenario where the original signal is sampled at this frequency and results in the same reconstruction). The rolloff of the filter is actually the result of the windowing, because the sinc time domain pulse shape that is used to reconstruct cannot be infinite, and therefore it is not a perfect rect shape brick wall filter in the frequency domain. The rolloff of the upsampled signal is of course identical. The fact that the impulse sinc side lobes in the frequency domain that also get multiplied with it have decreased in magnitude slightly by this point means that the rolloff x images x side lobes is less, so there is less spectral leakage outside of the desired band, though twice as large. This means that aliasing under the rolloff has less effect, as well as it being moved outside the audible range.
The point of upsampling is to move the reconstruction/anti-aliasing filter well above audible frequencies, where it can be more gentle and not affect the audible spectrum. A high-quality filter at 22khz is hard to make in hardware without distorting phase and without cutting into frequencies below 20khz. A gentle filter at 88khz is much easier, and it can distort there all it wants without affecting the audible range. It's a simple engineering solution to a problem that could otherwise become audible... at least to some of us (not to me, not for a while :()