New SIMD arch

The idea is to keep the two-speed architecture, to still avoid PCLMULQDQ but to switch from one branch per structural char to one branch per chunk.

• POC
• bench
• SWAR version
• avx2 etc. since it becomes potentially viable
• try to make the parsing more robust to quotes that are not in beginning position

The pseudocode idea is the following:

per chunk, compute the newline + quote mask and the delimiter one
we mask the delimiter one with the structural one (we need to find the first set, then merge masks from here)

we process the delimiter mask in a loop
we get first structural from the other mask and we branch accordingly
  either continue
  either newline
  either quote

If we have quote, we consume remaining bits of the mask in hope of finding a next quote, in which case we go to Quote instead of Quoted, so we don't lose work.
We might even want to attempt returning in Unquoted in a single chunk

Start by locating the current two-speed parser and its benchmark setup, then compare the issue's per-chunk mask approach with the existing structural-character branching. Work through the POC, benchmark, SWAR, and AVX2 checklist, and verify that quote handling also covers quotes that are not at the beginning of a chunk.