[langchain] Classify agent chains by name on on_chain_start

[!NOTE]
This issue was generated with AI assistance and requires investigation and confirmation prior to being worked on.

Part of #543.

In on_chain_start, agent classification currently requires explicit metadata keys (otel_agent_span, agent_name, agent_type). Standard LangChain agent chains (e.g. AgentExecutor) without explicit metadata are either misclassified as workflows or suppressed.

OpenInference classifies any chain whose name contains "agent" as an agent span.

• Classify chains with "agent" in the chain/runnable name as invoke_agent (gen_ai.operation.name: "invoke_agent")
• Resolve agent name from chain/runnable name
• Set standard agent attributes (gen_ai.agent.name, gen_ai.conversation.id, gen_ai.input.messages, gen_ai.output.messages)

Start by locating the on_chain_start entry point and the existing classification logic for otel_agent_span, agent_name, and agent_type. Compare its behavior with standard LangChain chains such as AgentExecutor, then verify that names containing "agent" produce invoke_agent spans with the requested agent name and standard attributes; add or update focused tests if the surrounding code provides them.