AVX2

This PR (#27402) introduces AVX2-specific kernel optimizations to significantly accelerate prompt processing and perplexity (PPL) calculations for IQ (Importance Quantization) models on CPU backends. ▶ Significant Throughput Gains: Benchmarks on high-end hardware, such as the EPYC 9654, demonstrate substantial speedups for Qwen-series models (27B, 35B-A3B) during large batch operations. ▶ Full Spectrum Support: The optimization covers the entire IQ range, from ultra-low bit (IQ1_S) to standard (IQ4_NL), ensuring robust performance across various tensor types. ▶ Workflow Acceleration: By streamlining iMatrix and PPL calculations, this update drastically reduces the latency associated with model evaluation and quantization tuning on local hardware. Bagua Insight In the Local LLM ecosystem, Importance Quantization (IQ) is the gold standard for maintaining high perplexity at ultra-low bitrates, yet its computational overhead on CPUs has historically been a bottleneck. This AVX2 optimization by bartowski1182 is a strategic play to rebalance the trade-off between precision and speed at the instruction-set level. As enterprise RAG pipelines increasingly demand efficient long-context processing, these micro-optimizations for large batch sizes are essential for validating CPU-based inference. It signals a shift where high-performance CPUs are no longer just a fallback for GPUs but are becoming viable engines for large-scale model evaluation and specialized inference tasks. This is a critical step in pushing the boundaries of edge and on-premise AI capabilities. Actionable Advice Developers relying on CPU backends for model quantization and benchmarking should immediately integrate and test this PR, especially when running iMatrix weight generation, where AVX2 will yield the most significant time savings. For enterprise architects, this update warrants a re-evaluation of IQ-quantized models for production; the combination of low memory footprint and improved CPU latency makes IQ models a much stronger candidate for high-throughput, non-GPU deployment scenarios.