Add POWER8 LLM inference and NUMA weight banking projects - #161

Adds two CPU inference projects to the CPU/Single GPU/FPGA/NPU/Mobile Inference section:

• llama-cpp-power8 — POWER8 VSX/AltiVec optimizations for llama.cpp featuring vec_perm non-bijunctive attention collapse, IBM MASS math library integration, and dcbt resident prefetch hints. Achieves 8.8x speedup over stock llama.cpp on POWER8 S824.
• RAM Coffers — NUMA-aware weight banking system for LLM inference on multi-socket servers. Routes model layers to specific NUMA nodes based on cognitive function mapping, with selective prefetch and non-bijunctive prune-before-fetch.

Both target CPU-only inference on non-x86 architectures (IBM POWER8), which is underrepresented in the current list.