Maybe (I'm not sure) because into_iter() actually copies i32, and iter() creates references. While references are also 32 bit on x32 arch, they are 64 bit on x64 arch. So it may cause unnecessary creation of muliple u64 pointers that slow down the whole thing.

using iter() instead of into_iter() will degrade performance

This is really weird.

Isn't into_iter giving you a copy, while iter giving you a reference? Why is making a copy faster?

If it were only the sum() or only the fold() as on stackoverflow this would be true.
But benchmarking shows that combining map() and sum() is about 10% slower.
Also, using iter() instead of into_iter() will degrade performance by another 30-40%.

Using vectors with 100 million elements on my machine:
iter + map + sum = ~3.1-3.3s
into_iter +map + sum = ~2.5-2.6s
iter + fold = ~2.9-3.1s
into_iter + fold = ~2.3-2.4s

Actually IntStream.of() calls Arrays.stream()
From IntStream.java :

<...>
/**
 * Returns a sequential ordered stream whose elements are the specified values.
 *
 * @param values the elements of the new stream
 * @return the new stream
 */
public static IntStream of(int... values) {
    return Arrays.stream(values);
}
<...>