According to the support article it is applied whenever the bit depth is reduced, so we would assume that may apply universally but at this time have no specific or definitive information if it applies from 32-bit float to 24-bit.
- Use Dithering for Audio Devices and Audio File Export is engaged by default, and means that triangular dithering (with no noise-shaping) is applied when the audio signal’s bit depth is reduced from a higher bit depth by a device or during file export. Turn this off if you would like to use a third-party dithering solution, such as a limiter Insert effect on the main output that has built- in dithering with characteristics you prefer.
To your other question, converting to 32-bit float (when bouncing) and then back to 24-bit later should not result in any loss of quality. Bouncing to 32-bit float is largely to remove any remote possibility of a file clipping during a direct bounce.