[util-genai] Investigate streaming helpers for agent and workflow invocations

SyncStreamWrapper and AsyncStreamWrapper in opentelemetry-util-genai are designed around LLM inference invocations (chat, generate_content). When passed an invocation, they mark it as streamed and record client token streaming metrics (gen_ai.client.operation.time_to_first_chunk and time_per_output_chunk) on each chunk.

Agent (invoke_agent) and workflow (invoke_workflow) operations do not define these client token metrics under GenAI semantic conventions. As a result, agent instrumentations either avoid passing the invocation to super().__init__ (e.g. qwen-agent) to manage the lifecycle manually, or pass it and emit invalid metrics (e.g. agno).

Agent instrumentations also duplicate logic for accumulating stream chunks into output_messages based on content capture settings, and serializing structured content (Pydantic models, dataclasses) into JSON strings.

Investigate possible approaches in util-genai to:

• Support streaming agent and workflow invocations without recording client token timing metrics.
• Reduce boilerplate across agent instrumentations for chunk content accumulation and structured object serialization.

Start by reading SyncStreamWrapper and AsyncStreamWrapper in opentelemetry-util-genai, then compare the qwen-agent and agno instrumentation approaches described in the issue. Investigate how agent and workflow invocations can avoid unsupported client token metrics while sharing chunk accumulation and structured-object serialization. Done means a clear, project-consistent approach is identified and its expected behavior is documented.