An Argument For PCLMULQDQ

Love this project! I am still a beginner when it comes to Rust, but I noticed in the README that you chose not to use the classic pclmulqdq trick from simdjson. That inspired me to create my own minimal CSV parser. It shares a very similar interface, but very different internals. It uses the pclmulqdq trick to great effect (60% faster in some cases, if my benchmarking code is correct), though implementing it similarly in simd-csv would likely require a lot of work. The gains also vary a lot by target, with the most clear gain only on x86_64 targets. Would you be open to reviewing a PR for similar approach in simd-csv? I think using target-feature would potentially work as a way to make it available for targets that do support pclmulqdq without sacrificing portability.

Start with the README's discussion of why the classic pclmulqdq approach was not chosen, then compare the proposed approach in the linked csimdv-rs project with simd-csv. Review how target-feature could preserve portability and verify the claimed gains with the existing benchmarking approach; done means a concrete, reviewed direction for x86_64 and other targets.