Don't clone input buffer inside html5ever benchmark loop

[simonwuelker]

This clone operation should not be counted towards the benchmark time. When the benchmark was written criterion did not support per-iteration setup, but now it does using Bencher::iter_batched. Using this method reduces the benchmark time by 9% in lipsum.html.

Additionally, the way html5ever was used inside the benchmarks was wrong if there were <script> elements in the input.
To fully tokenize input in html5ever, you need to spin on the Tokenizer::feed method until it returns Done.
None of our benchmarks contain any script elements, but it's better to fix it anyways.