docs: map public eval datasets to voice boundary

[ai-hpc]

Summary

• add evaluation data usability guidance for public datasets
• separate voice-runtime audio/STT datasets from GenieClaw BFCL/home-state datasets
• replace GeniePod Home wording with NVIDIA Jetson Orin 8GB positioning
• ignore local generated eval data under tests/eval/local

Real Behavior Proof

• cargo fmt --all -- --check
• cargo test --all-targets
• cargo clippy --all-targets -- -D warnings
• grep -R "GeniePod Home" -n /home/speedy/genie-voice-runtime --exclude-dir=.git --exclude-dir=target 2>/dev/null || true

Jetson Validation

• Not run on physical Jetson; docs/config-only change, no runtime behavior changed.