Change the non-allocating benchmarks to allocating.

Previously we had several non-allocating benchmarks.
They were not representative to the most common use case where we want
to save all the decoded input, not just a fraction.
This patch updates the benchmark to save the decoded input in lists.