docs: publish Qwen3.8 27B 5070 Ti recipe

[feveromo]

Publishes the measured local-inference recipe from the Aug 19 optimization session.

Kept it self-contained under recipes/qwen3.8-27b-5070ti/ so the Pi Chisel project docs stay clean.

Includes:

• the actual 5070 Ti / 9800X3D / 32 GB setup
• controlled before/after numbers
• the auto-fit bottleneck
• embedded MTP n=2 config
• adaptive VRAM launcher
• systemd example
• benchmark script
• the stuff that lost: Q8 KV, bigger batches, Q2_K, Q4_K, NVFP4, DFlash2