Add local trace replay regression harness

[Ritwij Aryan Parmar (RitwijParmar)]

Summary

This adds a local trace replay path for the Python SDK. The goal is to make saved Braintrust trace exports useful as regression cases when iterating on an agent/task or scorer, without creating a new experiment just to sanity-check behavior.

What changed:

• added braintrust replay for JSON/JSONL span exports
• added ReplayTrace so replayed scorers can inspect spans with get_spans() and get_thread()-style access
• reports current scores, baseline root-span scores, score deltas, derived trace metrics, and metric deltas
• added CI-oriented gates: --min-score, --min-score-delta, and --fail-on-error
• documented the workflow in the Python README

Why

For agent/eval workflows, production traces often capture the hard cases: tool-call paths, bad intermediate states, or regressions that unit fixtures miss. This gives users a lightweight way to replay those traces locally and fail a check when a task or scorer change regresses against the saved baseline.

Tests

• PYTHONPATH=py/src .venv/bin/python -m pytest py/src/braintrust/test_trace_replay.py -q
• PYTHONPATH=py/src .venv/bin/python -m pytest py/src/braintrust/cli/test_push.py py/src/braintrust/test_trace_replay.py -q
• PYTHONPATH=py/src .venv/bin/python -m compileall -q py/src/braintrust/trace_replay.py py/src/braintrust/test_trace_replay.py py/src/braintrust/cli/__main__.py py/src/braintrust/__init__.py
• git diff --check