[DOCS] document the daemon-hosted A2A server, its a2a.server config and agents.yaml headers #1008

Summary

The CLI adds user-facing behavior in inference-gateway/cli#1556 (feat(a2a): host an A2A server in the daemon, one headless run per task) that the docs site does not cover yet:

• New config section a2a.server with enabled, host, port, token and mode keys plus INFER_A2A_SERVER_* environment overrides. Defaults are off, 127.0.0.1, 8090, standard mode. The server is off by default and the daemon refuses to boot when it is enabled without a token.
• infer daemon hosts an ADK A2A server: the agent card at /.well-known/agent-card.json, bearer auth, an in-memory task store and a one-at-a-time queue. Each incoming task runs as one headless run; a shared contextId continues the same session across tasks; tasks/cancel kills the run.
• New headless flag --prompt-only, which treats task text as a plain prompt (no ! or !! direct calls, no / shortcuts). The A2A handler always sets it.
• agents.yaml peer entries accept a headers map whose values are expanded with os.ExpandEnv, for example Authorization: Bearer ${PEER_TOKEN}, so one infer can authenticate to another.

The repository docs (docs/daemon.md, docs/agents-configuration.md and the README config table) already describe this. The docs site needs the same coverage on its affected pages.

Acceptance Criteria

• The configuration reference lists the a2a.server keys, their defaults, the INFER_A2A_SERVER_* overrides and the token requirement for boot.
• The daemon page documents the A2A server subsystem: agent card, bearer auth, one-task-at-a-time queue, session continuation via contextId and cancellation.
• The A2A client docs cover the agents.yaml headers map with os.ExpandEnv expansion.