I understand what the dithering process does, but I'm not clear what the difference is between the different types available when you are bouncing a project for export.
Two of the three options use noise shaping, which is basic filtering of the additional dithering noise.
If you understand the effect of dithering, you will probably be familiar with its big drawback: the introduction of noise. While the noise should be randomly generated to get the best dithering results, the most significant reason we dither is actually to prevent rounding errors when lowering our resolution, e.g. by getting back from 24 bits (or even 32 bits) to e.g. 16 bits for CD production.
Noise can be more irritating on certain frequency bands. (6 kHz is the most famous frequency band for irritating noise, as far as I know.) Noise shaping is used to get rid of this, by actually not adding white noise to the signal, but slightly filtered noise, so that the total amount of energy still masks the rounding errors, while the dithering signal is more pleasant to listen to.
Especially if you are going to downsample from 96KHz (or 88,2KHz) to 44,1KHz, it is useful to get the noise into the high frequency bands, since most of them will be filtered away during the downsampling.
I pulled up the Logic manual to take a look at what they said about this. Below is an exact copy from the book: