Handle environment manager capacity limits

[zhenyusteven]

What changed

• Allow the environment manager and spawned environment servers to bind through SERVER_BIND_HOST while preserving the existing IPv6 default.
• Return HTTP 503 when environment capacity is exhausted instead of surfacing an unstructured internal error.

Why

Large asynchronous RL rollouts can temporarily consume every configured environment slot. Callers need a retryable capacity response, and some cluster deployments require an explicit bind host.

Impact

Existing deployments retain the :: bind default. Capacity-aware clients can now distinguish temporary resource exhaustion and retry safely.

Validation

• python3 -m py_compile online_server/online_server/env_instance_manager.py
• git diff --check