Polish the personal single-GPU release - #13

Summary

• redesign the banner and benchmark figure with a dark visual system and non-overlapping axes
• make PyPI and the bring-your-own-GPU path prominent in both READMEs
• generalize the supported recipe with automatic bf16/fp16 selection while preserving measured RTX 4080 evidence
• prepare the v0.2.2 release metadata and regression gates

Validation

• 1066 non-GPU/non-network tests passed; 87.11% coverage
• 5 GPU tests passed on RTX 4080
• 3 network tests passed
• ruff, formatting, mypy, package build and twine checks passed
• frozen benchmark JSON SHA-256 remains 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc