At the risk of being trite, it's down to the order of evaluation.
You know that a for comprehension expands into calls to flatMap and map? So
for {
a <- forA
b <- forB
} yield a + b
expands to
forA flatMap { a => forB map { b => a + b } }
Remember that forB creates and returns a future. Only when this future has been created will it start running.
It should be apparent that forB is not evaluated till forA's future delivers a value, and thus the future returned by forB will run in serial with the future returned by forA.
In the other example the futures are created outside of the for comprehension, and thus start running in parallel.
I can't see the difference between your examples and so am confused why 1.) is serial 2.) is parallel.