Both loops get entirely unrolled. Its 5 instructions for each iteration in the first example, and only 3 in the second example. (To say nothing of the fact that conditional jumps are (usually?) much more expensive than add/mov instructions)
I believe the compiler can't do that, since the bounds check panic is a side effect that can be observed - the message tells which index that failed the check! For that reason, non-elided checks will not be reordered.
However, I believe a future Rust RFC could turn that around and validate the idea that in some cases such things could change execution order, even if it has noticeable side effects.
In practice a small function like this would be inlined, which gives room for further optimisations. At any point, if the compiler knows that all accesses are in bounds, it can remove the bounds checks. The trick is actually having it figure that out.