feat(agent-runtime): add typed agent lifecycle

[nvddr]

Summary

• add the dependency-light xr-ai-agent-runtime package with one addressable Agent lifecycle
• expose call(ref, message) for bounded request/response and publish(topic, event) for typed fan-out
• let resident agents start runtime-owned background tasks while bounded mailbox handlers serialize private state
• migrate simple-vlm-example finite vision requests to an addressable SimpleVlmAgent

Architecture

Raw media stays on XR-Media-Hub. The runtime carries typed semantic messages with participant, correlation, source, and parent metadata. The dispatcher stays private; model loops, tools, planning, and memory remain agent implementation details. Relay remains the execution and telemetry boundary around model and tool calls inside agents. A future Fabric adapter can host or bridge this lifecycle without changing application-level call and publish topology.

The simple VLM worker now registers LiveVisionTool behind the runtime. Its bounded agent returns one complete VisionResponse; streaming voice behavior is intentionally layered in #351.

Review feedback addressed

• every call and subscriber receives a deep-isolated Pydantic payload
• a call timeout covers mailbox backpressure and handler execution as one operation
• participant scope is runtime metadata and is preserved across nested calls and publications
• runtime ownership replaces the former manual periodic-source lifecycle

Validation

• runtime and simple-VLM focused suite: 18 passed
• targeted Pyright: 0 errors
• Ruff, SPDX, DCO, worker wheel build, and worker lock resolution passed

Rebased onto main after #349 merged. #351 is stacked on this PR.

[wenxind-nvidia]

Blocker

1. Runtime-owned failures are silently downgraded to event-loop log messages. AgentContext._background_done() lets a failed background task disappear while runtime.running remains true, and AgentRuntime._deliver() suppresses subscriber exceptions so publish() returns successfully. I reproduced both behaviors: essential work can stop while the worker still appears healthy. Please let subscriber failures propagate after all fan-out deliveries have settled, and make a runtime-owned background failure fail or explicitly transition the runtime instead of only invoking the loop exception handler.

[nvddr]

@wenxind-nvidia Addressed in 182a4fc. Subscriber exceptions now propagate from publish() only after every fan-out delivery has settled. A failed runtime-owned background task now transitions the runtime to failed, sets running false, cancels remaining owned work, causes later operations to raise RuntimeFailedError from the original exception, and surfaces that original failure again during shutdown. Added regression coverage for both cases. The 28 focused runtime/tool tests pass on Python 3.11 and 3.12; Ruff and Pyright also pass.