Add custom trajectory helpers for external agents

[eldwin-easynet-world]

Summary

This adds a small custom-trajectory helper module for external agents that already produce BixBench answers/notebooks and need a safe way to validate and write postprocessing-compatible JSON before running a full benchmark.

Changes:

• add bixbench.custom_trajectories with validate_custom_trajectory, write_custom_trajectory, and minimal_notebook
• document a no-core-edit custom-agent flow in the README
• make load_dataframe_from_json_directory tolerate JSON files without a replica_<id> suffix by defaulting to replica 0
• add focused tests for validation, writer naming, and directory-loader compatibility

Why

The current README tells custom-agent users to edit custom_rollout and produce the same trajectory shape as BixBench trajectories. That is hard to smoke-test before spending credits or running Docker/Hugging Face/full agentic evals. This patch gives external agents a small contract and deterministic checks for the trajectory format used by postprocessing.

Validation

Ran in a local Python 3.12 conda environment:

PYTHONPATH=. conda run -n bixbench-py312 pytest tests/test_custom_trajectories.py -q
PYTHONPATH=. conda run -n bixbench-py312 python -m py_compile bixbench/custom_trajectories.py bixbench/postprocessing_utils.py tests/test_custom_trajectories.py
PYTHONPATH=. conda run -n bixbench-py312 pytest tests -q

Results:

• tests/test_custom_trajectories.py: 4 passed
• full local suite: 19 passed, with existing Pydantic v2 deprecation warnings