perf: Triton-accelerate full pipeline (quantize + dequantize + WHT rotation)

perf: Triton WHT rotation kernel for O(d log d) fused rotation

perf: wire Triton fused quantize kernel into GenerationCache

perf: wire Triton fused dequantize kernel into GenerationCache

perf: add speed benchmark for Triton vs Python quantize/dequantize

feat: overnight validation — 236B model running on single RTX 4090

feat: cross-layer KV cache with shared codebook/rotation resources

feat: 1-bit value quantization with correction (V compression is free)

feat: self-correcting KV cache with periodic refresh (prevents error accumulation)

feat: ultra-streaming engine for 200B+ models on consumer GPUs

perf: switch to WHT rotation as default (O(d log d), 98% quality match)

feat: unified 70B-on-4090 launcher with auto-configured KV compression

feat: integrate HybridCache + fixed eviction + streaming into autoresearch sweep

feat: TurboQuant weight compression (TQ-W) for ultra-low-bit model deployment

feat: HybridCache combining boundary anchoring + gradient bits + per-head allocation