It doesn't matter which nonces you test or in what order that other than you need to avoid testing the exact same block header twice (since that would obviously be a waste of time).

It's useful to test all the nonce values since if you don't test them all you will need to update extranonce more often, though not critical.

Even though miners increment their nonces in similar ways this doesn't result in the duplication of work between miners: there are a myriad number of differences between the candidate blocks miners attempt, foremost being where their reward is paid. (And pooled miners that happen to pay to the same place get different extranonce ranges assigned by the pool). As a result there is never any work duplication across miners unless something is embarrassingly broken.

Starting at 0 and incrementing until the maximum value is simply a convenient way to change the nonce that cannot duplicate and minimizes extranonce updates. Many devices assign distinct ranges to different chips, and so don't simply increment in a simple 0 to 2^32-1 manner.

A mining device could save a negligible amount of power by changing using a linear feedback shift register instead of an adder, but no one seems to have bothered with that optimization.