Expose response length on BenchmarkRun

[amargupta0428]

Expose response length on BenchmarkRun

Closes #1

What and why

Computing model response length currently requires manually downloading and
parsing the eval logs referenced by BenchmarkRun.logs (typically on S3) with
Inspect's read_eval_log. This PR adds two methods to BenchmarkRun so the
client library handles that directly:

• BenchmarkRun.read_eval_log(**kwargs) — fetches and parses the run's eval
  log from the URL in the logs field using Inspect's read_eval_log. Extra
  keyword arguments are passed through.
• BenchmarkRun.response_lengths() — returns the response length, in output
  tokens, for each sample in the run. Samples without recorded token usage
  (e.g. failed samples) count as 0.

inspect-ai is imported lazily, so it remains an optional dependency: it is
only required when these methods are called, and a clear error message is
raised if it is missing. Existing users who don't need eval logs are
unaffected.

Usage

from epochai.airtable.models import BenchmarkRun

runs = BenchmarkRun.all(memoize=True)

for run in runs:
    lengths = run.response_lengths()
    print(run.inspect_run_id, sum(lengths) / len(lengths) if lengths else None)

Implementation notes

• inspect_ai.log.read_eval_log resolves the logs URL itself, including
  s3:// paths (S3 credentials must be configured as usual for Inspect).
• No changes to existing fields or models; no new required dependencies.

Tests

Added tests/test_benchmark_run.py (plus tests/conftest.py providing dummy
Airtable env vars). Tests stub out the inspect_ai.log module, so the suite
runs fully offline:

.venv/bin/python -m pytest tests/ -v
# 4 passed