## Feature description

The DataDogLLMObs logger (`litellm/integrations/datadog/datadog_llm_obs.py`) ships LLM Observability spans over HTTP, with two configuration modes:

1. **TCP agent** (`LITELLM_DD_AGENT_HOST` set → `http://<host>:8126/api/intake/llm-obs/v1/trace/spans`) 2. **Direct intake** (`DD_API_KEY` + `DD_SITE` set → `https://api.<site>/api/intake/llm-obs/v1/trace/spans`)

Neither works when the Datadog Agent is reachable only via a **Unix domain socket**, which is the modern Datadog Operator default.

## Why this matters

Datadog Operator deployments using APM auto-instrumentation mount the agent socket at `/var/run/datadog/apm.socket` and set `DD_TRACE_AGENT_URL=unix:///var/run/datadog/apm.socket` on instrumented pods. The TCP receiver on port 8126 is **not** bound on the host network in this configuration — see the upstream agent docs on UDS receivers.

The `ddtrace` library handles this transparently (it consumes `DD_TRACE_AGENT_URL`). But the LiteLLM `datadog_llm_observability` callback hard-fails because:

- TCP agent mode resolves `<LITELLM_DD_AGENT_HOST>:8126` and gets `ConnectionRefusedError`/`ClientConnectorError` on every flush. - Users have to fall back to di…