docs: turbo4 4.125 bpw + Jun-2026 rematch writeup; pin 3/4-bit centroid values

[TheTom]

Research-repo counterpart to llama-cpp-turboquant#197. Factual fixes + a writeup + a regression test, no algorithm change (the Python reference was already correct).

Changes

• README: turbo4 is 4.125 bpw / 3.9×, not 4.25 / 3.8× — the no-QJL 4-bit block is norm(fp16) + 64 B indices = 66 B; the old number counted a phantom rnorm. (The C/CUDA port also dropped that dead field.)
• docs/turbo4-rematch-2026-06.md: the June-2026 turbo4 work + head-to-head vs spiritbuun/master (KLD/prefill/decode at 4.125 bpw), the centroid-port-bug story, PDL backport + fused-MMA decode, and the asymmetric q8_0-K + turbo4-V result (−26% KLD; "V is free" holds at 4-bit, breaks at 2-bit).
• tests/test_codebook.py: pin the exact 3-bit and 4-bit optimal centroids for d=128. The suite previously pinned 1-bit/2-bit but only sanity-checked 3/4-bit — exactly the gap that let a downstream C/CUDA port drift to a mis-scaled 4-bit table (outer 0.1739 vs correct 0.2402, ~2.1× excess MSE). Now any port can be checked against these.

Notes

• The Python reference codebook was correct throughout; this is documentation + a guard, not a fix to optimal_centroids.
• pytest tests/test_codebook.py → 26 passed.