Implement faster tokenizer versions for `DOMParser`/`innerHTML`

The main tokenizer needs to count line numbers, which introduces significant overhead, especially in the SIMD case.

These line numbers are not required in some situations and we should add an option (perhaps with a `bool` const generic?) to disable them.

Gecko has something similar in https://searchfox.org/firefox-main/source/parser/html/nsHtml5TokenizerLoopPoliciesSIMD.h.