Retry once on stale-connection errors from shared aiodocker client after dockerd restart

[rapsealk]

Parent epic: #11216
Follow-up to: #11226

Main idea

#11226 pooled a single long-lived aiodocker.Docker (wrapping an aiohttp.ClientSession) on DockerAgent, replacing the per-op async with closing_async(Docker()) as docker: pattern. That's a big latency win, but it introduces one production failure mode that the old pattern didn't have:

After systemctl restart docker (or any dockerd bounce), the agent's long-lived aiohttp.ClientSession still holds keepalive sockets in its connector pool. The first call to dockerd after the restart picks a stale socket, fails with aiohttp.ClientConnectionError or ServerDisconnectedError, and surfaces as a spurious failure in purge_images, scan_images, check_image, or any per-op site. Aiohttp reconnects on the next call, so the error is one-shot — but that one shot is enough to fail a user-visible operation.

Per-op Docker() never hit this because each call opened a fresh socket.

Design

Add a thin once-retry wrapper around the per-op aiodocker sites that catches aiohttp.ClientConnectionError / aiohttp.ServerDisconnectedError exactly once, then re-runs the call. On the second failure, propagate normally. No exponential backoff — the retry is specifically for the stale-connection case, which resolves on the next attempt.

Sites to cover (grep self.docker. in src/ai/backend/agent/docker/agent.py):

• apply_accelerator_allocation
• start_container
• get_intrinsic_mounts
• destroy_kernel / clean_kernel
• extract_image_command
• enumerate_containers
• scan_images
• push_images / pull_images / purge_images
• check_image
• create_local_network / destroy_local_network
• resolve_image_distro

Do NOT wrap:

• monitor_docker_events() — it already has its own reconnect loop over closing_async(Docker()).
• Streaming reads (DockerStatsStreamer, introduced in refactor(BA-5859): Stream container stats from Docker instead of polling #11224) — those already have bounded-backoff reconnect.

Alternative ideas

• Set a short force_close=True on the connector so every request uses a fresh socket. Simpler, but negates refactor(BA-5858): Reuse a long-lived aiodocker client across container operations #11226's latency win. Rejected.
• Health-probe on interval and recreate the client when dockerd version endpoint fails. More infrastructure, catches the same case. Overkill for now.

Out of scope

• Accelerator plugins' Docker() usage — tracked separately within epic Improve agent container resource-mounting performance (aiodocker + runtime layer) #11216.

JIRA Issue: BA-5862