1) dont think Tesla uses a complex ML algorithm to make predictions as from what I have gathered, they use some variant of (your num minus current num) divide by daily production, type models.
2) If they were to use an ML model with an ordinal output, they can either sum all the days together and provide a range based on the 95% confidence intervals derived from standard errors, or run a bunch of simulations and provide a range based on the 95% confidence interval of the simulation outputs. Again, if it were me, I would not use an ML model for this problem. Maybe for some tiny component, but definitely not for the main part of the algorithm as the EDD is still largely dependent on how far back a person is in the queue divided by the daily production counts.