Add KV-cache key compression reproducible comparison

[yi-fireworks]

Summary

Self-contained kernel-level comparison of three KV-cache key compression methods
(Conventional scalar quantization, TurboQuant, RotorQuant) on a B200 GPU with
matched storage budgets and fused Triton scoring kernels.

• Measurement harness, library code, and reference results at research/kv-cache-compression/
• Covers 2k/4k/32k context, 8/4/3/2 bit widths, quality and latency
• README documents setup, reproduction, and interpretation

Directory structure

research/kv-cache-compression/
  README.md                    # Standalone entry point
  requirements.txt             # torch, triton, scipy, transformers
  configs/kernel_table_phase1.json
  scripts/run_unified_table.py # Measurement harness
  turboquant/                  # Library (9 .py files)
  results/unified/             # Reference CSV + markdown

What was adjusted from the internal experimental branch

• Removed stale baseline block from config JSON (referenced internal paths
  and an outdated note about RotorQuant being over-budget)
• Added transformers to requirements.txt (needed by an import in fused_attention.py)
• Added .gitignore
• Excluded: __pycache__/, patch_test.py (empty stub), data/LLMTest_NeedleInAHaystack/
  (vendored third-party, unused by harness), results/niah/ and results/real_keys_eval/
  (separate experiments not documented in README)

Reproduce

cd research/kv-cache-compression
pip install -r requirements.txt
python scripts/run_unified_table.py --gpu 0

Requires any Ampere+ NVIDIA GPU (A100, H100, L4, B200, etc.). Quality numbers
are seed-deterministic and should match the reference CSV. Timing varies by hardware.

Test plan

• All Python files pass py_compile
• Core import chain resolves (TurboQuantProd, RotorQuantMSE, LloydMaxCodebook)
• No internal paths, credentials, or infra references (verified via ripgrep)
• Reference CSV well-formed (87 rows, 15 columns)
• GPU smoke test when hardware access returns