To explain the supposed contradiction between the number of samples calculated by symbol length and sampling rate on the one hand and the IFFT size on the other hand, you have to take into account that the IFFT output is upsampled by a factor of 4. This means that for every output sample of the IFFT block there are four samples after upsampling. There's a cyclic prefix of 16 samples so that we arrive at a total number of samples of $N=4\cdot(64+16)=320$ as stated on page 53.
On the other hand we can calculate the number of samples by $N=Tf_\mathrm{S}=32\mathrm{ms}\cdot8\mathrm{kHz}=256$. The symbol duration $T$ has been taken from page 108. It think the guard interval has not been taken into account there and thus $N=64\cdot4=256$.
The process that is described as "upconversion" is exactly the frequency shift that you are observing (cf. figure 3-8). Citing from the report:
Once the baseband signal has been interpolated it is multiplied with
the audio band carrier frequency to obtain a audioband representation.
To make this more clear using a formula, let $x(t)$ be the complex, analog OFDM signal (after digital-to-analog conversion (DAC)) and let $f_0$ be the desired carrier frequency. Then the radio frequency, upconverted signal $y(t)$ is
$$
y(t) = x(t)\operatorname{exp}(j2\pi f_0 t)
$$
This is a basic concept required in almost every radio system. For more information you might want to do some reserach on radio frequency upconversion and I/Q modulation.
The baseband bandwidth of an OFDM signal is approximately $f_\mathrm{S}/2$. But again, here we have to take the upsampling into account. If it is done correctly, it won't change the bandwidth of the signal. I.e. the sampling frequency before upsampling is actually $f_\mathrm{S}/4 = 2\mathrm{kHz}$ and we have to use this frequency for calcuating the OFDM signal bandwidth wich here is $B=1\mathrm{kHz}$. Finally, the band occupied by the radio frequency signal is $f_0 - B\ldots f_0+B$ which corresponds to 1kHz to 3kHz.
Concerning your suggestion to use a greater IFFT size and to left some subcarriers unmodulated to achieve upconversion - this is a valid approach in this specific example. Note however, that in general it requires a larger sampling frequency which is often undesirable.
