TurboQuant KV cache port to v2

[jrajala6]

Summary

Engine KV cache is now TurboQuant unconditionally with no INT8 path. Default config: K=6, V=4, single Hadamard, seed=42. Graph-level INT8 primitives stay in cactus-graph (still exercised by test_cache) but unreachable from the engine.

Headline: 43% smaller cache vs INT8, 1.6% faster decode, ≤5% perplexity vs INT8 on 8 of 9 measured cells.

Changes

• cactus-engine/{src/engine.h, src/model.cpp, models/gemma4/model_gemma4.cpp} — TQ unconditional; INT8 fallback removed.
• cactus-kernels/cactus_kernels.h — TURBOQUANT_ROTATION_LAYERS 3 → 1 (single Hadamard; 3× cheaper rotation, equal/better quality per prior data).
• cactus-engine/tests/{test_perplexity_context_sweep.cpp, test_decode_latency.cpp} — new. Multi-corpus / multi-seed / multi-ctx sweep + real decode benchmark.
• cactus-engine/tests/test_perplexity.cpp — drops INT8 column.
• python/src/tensor_io.py — restores INTERLEAVE_BLOCK = 4 (upstream b99db766 removed the constant but left two functions referencing it in default args, breaking module import).

Validation — gemma-4-e2b vs INT8 cache (|TQ/INT8 − 1| ≤ 0.05)

INT8 reference numbers were captured during development against the in-tree INT8 cache before this PR removed it; reproducible from the prior commit.

Quality at seed=42 across context lengths:

corpus	ctx=1024	ctx=2048	ctx=8192
wikitext	−3.9% ✓	−9.1% ✗	−5.1% (borderline)
code	+4.6% ✓	+4.4% ✓	+2.8% ✓
chat	−2.5% ✓	–	–
math	−4.0% ✓	–	–

Seed robustness (K6V4, ctx=1024):

corpus	seed=42	seed=7	seed=1337
wikitext	−3.9% ✓	+7.6% ✗	−7.0% ✗
code	+4.6% ✓	+1.2% ✓	−1.5% ✓
chat	−2.5% ✓	+5.7% ✗	−17.2% ✗
math	−4.0% ✓	+4.4% ✓	+0.7% ✓

seed=42 is the only seed passing on every corpus → ship pinned to 42; other seeds remain reachable via CACTUS_KV_TQ_SEED.

Decode latency (gen=64): INT8 = 115.48 ms/tok, TQ K6V4 = 113.66 ms/tok (−1.6%, faster).

Memory (head_dim=256): FP16 1024 / INT8 576 / K6V4 328 bytes per token. K6V4 is 3.12× vs FP16, −43% vs INT8. At ctx=32k across 15 own-KV layers: 276 MB → 158 MB (saves 118 MB).

Why K6V4 over alternatives

config	worst-case 4-corpus	memory vs INT8	verdict
K8V8	2.0%	−10%	safe but small win
K6V4	4.6%	−43%	shipped
K4V4	+22.9% (wikitext)	−54%	fails gate
K4V2	+49% (code), ±75% seed swing	−64%	fails gate, seed-unstable

Configurability

CACTUS_KV_TQ_K_BITS, CACTUS_KV_TQ_V_BITS, CACTUS_KV_TQ_SEED, CACTUS_KV_TQ_BITS (unified). CACTUS_KV_WINDOW_SIZE, CACTUS_KV_SINK_SIZE retained from prior config.

Test plan

cmake -DCACTUS_BUILD_TESTS=ON -S cactus -B build && cmake --build build -j
./build/cactus-engine/cactus-graph/cactus-kernels/test_turboquant_smoke
./build/cactus-engine/cactus-graph/test_cache

CACTUS_TEST_MODEL=<weights/gemma-4-e2b-it> \
CACTUS_PPL_CORPUS=<corpus.txt> CACTUS_CTX_SWEEP=1024 \
CACTUS_TQ_SEEDS=42 CACTUS_SWEEP_CONFIGS=K8V8,K6V4 \
./build/cactus-engine/test_perplexity_context_sweep

CACTUS_TEST_MODEL=<weights/gemma-4-e2b-it> \
CACTUS_DECODE_PROMPT_LEN=512 CACTUS_DECODE_GEN_TOKENS=64 \
./build/cactus-engine/test_decode_latency