perf: cut CLI startup and batch overhead

[Microck]

Summary

• cache shared reqwest::Client instances by timeout instead of rebuilding them on hot paths
• keep batch search responses structured until the final output boundary instead of serializing and parsing them twice
• switch the CLI entry runtime to Tokio current_thread to reduce short-invocation overhead

Measured Impact

• composite benchmark baseline: 1152.092 ms
• retained-branch best run: 528.389 ms (-54.1%)
• retained-branch final verification run: 581.169 ms (-49.6%)
• note: the harness includes real networked Kagi calls, so news/search/quick/batch retain some upstream variance; startup-only commands improved more consistently

Verification

• cargo build --release && ./.tmp/benchmark-metrics.py
• cargo test --quiet

Notes

• autoresearch logs and benchmark artifacts were kept uncommitted
• discarded experiments included parser selector caching, lazy Tokio init for async-only commands, and release-profile LTO/strip tuning after it exhausted local disk during linking