Speed up zip_broadcast() by pre-filling the scalar elements

[kalekundert]

#737 got me thinking about how to speed up zip_broadcast(), and one simple optimization I noticed is that there's no need to initialize the new_item list and fill in all the scalar values on every iteration. That only needs to be done once up front, since the loop will only ever alter the positions in new_item associated with the iterable arguments.

For the example given in #737, this new version is twice as fast as the old version:

New version:

>>> timeit('consume(zip_broadcast("_", "__", [1] * 100_000, [222] * 200_000))', globals=globals(), number=10)
0.45314044202677906

Old version:

>>> timeit('consume(zip_broadcast("_", "__", [1] * 100_000, [222] * 200_000))', globals=globals(), number=10)
1.0282995470333844