Add eval-as-a-gate scoring

[Siddhant-K-code]

Summary

• add named raw-value eval criteria for cost, errors, completion status, redundant reads, tool calls, context use, duration, and lint findings
• support versioned baselines, regression tolerance, JSON/table output, custom Python scorers, and GitHub Actions step summaries
• preserve the legacy eval subcommands while making agent-strace eval [SESSION_ID] the direct CI gate
• wire the existing composite action to the named-gate workflow safely
• document .agent-evals.yaml and bump agent-strace from 0.83.0 to 0.84.0

Closes #210

Verification

• PYTHONPATH=src python -m unittest tests.test_eval_gate tests.test_eval tests.test_eval_extensions -q — 96 tests passed
• Ona test task — 1,680 Python tests and 3 VS Code extension tests passed
• python -m compileall -q src/agent_trace && git diff --check
• wheel build — agent_strace-0.84.0-py3-none-any.whl

[Siddhant-K-code]

PR review completed: checked backward compatibility, direct/legacy CLI routing, baseline direction and tolerance semantics, GitHub Actions argument handling, storage compatibility, and coverage. No blocking findings. Merge remains gated on all five GitHub CI jobs.