Tip: use this llama.cpp PR to improve PP on Intel ARC

A community pull request for llama.cpp significantly improves prompt processing speed for Intel ARC users, specifically benefiting hardware like the B580. The contributor optimized the code with assistance from Claude to accelerate context handling.

• Processing a 116k context conversation dropped from 510 seconds (245t/s) to 262 seconds (462t/s) using Qwen3.6 35B A3B Q5_K_XL.
• The optimization currently supports F16 KV cache, with plans to extend support to other quantizations later.

This improvement brings Intel ARC hardware closer to its full potential through continued community contributions.