Dice rolls are optional and are not required to address this issue.
I'd ignore that "advice" and do them regardless. Even though the RNG gives 32 bits worth of data at a time, it doesn't mean you get 32 bits of actual entropy; any bias in the RNG weakens entropy but not the length of the data it returns. This is tricky to detect.
With dice rolls, preferably with a set not a single one, 99 times, you basically add a whole 256 bits of random that is completely separate from any hardware RNG. More separate sources of entropy = better, as long as the combination mechanism is hardened (i.e. stream it through sha256, which is what coldcard does.)
I'd ignore that "advice" and do them regardless. Even though the RNG gives 32 bits worth of data at a time, it doesn't mean you get 32 bits of actual entropy; any bias in the RNG weakens entropy but not the length of the data it returns. This is tricky to detect.
With dice rolls, preferably with a set not a single one, 99 times, you basically add a whole 256 bits of random that is completely separate from any hardware RNG. More separate sources of entropy = better, as long as the combination mechanism is hardened (i.e. stream it through sha256, which is what coldcard does.)