Furthermore, sleep is very specifically about the taking the brain offline: that's what deteriorates first in the absence of sleep and the tortured workarounds for animals that absolutely must avoid sleep (e.g. migratory birds) involve sleeping part of the brain at a time. Any explanation that isn't highly specific to the brain's responsibilities has the immediate hurdle of explaining this away, and for that reason I don't buy the mitochondrial explanation. Mitochondria are too universal and sleep is too specific to the brain. Energy is fungible, so I don't buy that nature wouldn't figure out the "trick" of having a subset of the mitochondrial population sleep at a time.
My money is on the "brain algorithm" requiring an online/offline phase, whether that's contrastive learning or memory consolidation or something else. There are lots of candidates for fundamental brain algorithms with the "feature" that they require an offline phase that cannot be incrementally worked around, and these fit the observations much better.