Don't compute ((uint64_t)1)<<64 in round_to_power_of_2

This would be undefined behavior if it happened. (It can't actually
happen as we're using round_to_power_of_2, since we would have to
be trying to allocate exabytes of data.)

While we're at it, fix the behavior of round_to_power_of_2(0),
and document the function better.

Fix for bug 6831.