You can see how resampling can change, and especially, increase the amplitude of a signal, as well as adding intermediate amplitude values. Since integer (8/16/24 bits) formats will clip any signal beyond 0dBFS, you need to convert to floating point BEFORE resampling to prevent this to happen. If a specific program does not support float, then you need to have some speculation about potential clipping, and apply a negative gain to the signal prior to resampling. If you do this, you still need to increase the signal to a higher integer bit-depth to minimize quantization error BEFORE applying a negative gain.
For 16/44.1 to 32/48, if the 32 is float, there is no way to dither because dithering relies on fixed quantization steps, but floating point formats don't use fixed quantization steps. If the 32 is integer, some software may still allow dithering, but not all software offer this option as it is unimportant anyway, the quantization error in this bit-depth is too small to begin with.