Handle stale connections in fetch_base_experiment()

[dpdiliberto]

Problem

Long eval jobs (500+ docs) on Azure VMs freeze for ~15 minutes after reaching 100% completion. The failure occurs when summarize() calls fetch_base_experiment(), which uses app_conn() (Vercel IP) that sits idle during the eval run.

Azure NAT gateways have a 4-minute idle timeout that silently closes stale connections. When fetch_base_experiment() tries to use the stale connection, it fails with ConnectionError.

Solution

Added retry logic with connection reset to fetch_base_experiment():

• 3 retry attempts with exponential backoff
• Explicit timeouts (5s connect, 10s read)
• Connection reset (conn._reset()) on retry to create fresh HTTP session
• Returns None after max retries instead of raising

Changes

• logger.py: Added retry logic to fetch_base_experiment()
• test_stale_connection.py: Integration test using real HTTP server to simulate NAT timeout

Testing

cd sdk/py
PYTHONPATH=src python3 -m unittest braintrust.test_stale_connection -v

Test uses a real HTTP server that simulates NAT gateway timeout behavior (0.5s timeout simulates 4-minute Azure NAT timeout).