feat(benchmarks): add agent-oriented math and transform benchmark cases

Summary

Adds three small agent-oriented benchmark cases under \enchmarks/agent/\ exercising patterns that code agents commonly generate or repair, as requested in issue #8.

Cases

• gcd: Euclidean GCD with integer loop and branch behavior
• stats: Min, max, and sum over a fixed-size array
• token-count: Delimiter token scan over a byte span (space-delimited)

Changes

• \enchmarks/agent/gcd.0\ - GCD benchmark source
• \enchmarks/agent/stats.0\ - Stats benchmark source
• \enchmarks/agent/token-count.0\ - Token count benchmark source
• \scripts/bench.mts\ - Registered new cases in benchmark harness
• \docs/articles/benchmarks.md\ - Documented agent-oriented cases section

Verification

These files follow the same syntax patterns as existing \enchmarks/zero/\ and \enchmarks/rosetta/\ cases but have not been locally built/tested (Windows host without the new-syntax-compatible build). Recommended to verify locally with:

sh pnpm run bench

Closes #8