langchain: replace direct callback-handler unit tests with tests that drive the LangChain API #820

Many langchain instrumentation tests call OpenTelemetryLangChainCallbackHandler.on_* methods directly with hand-built arguments, instead of running real LangChain components:

Problems:

• Bugs go unnoticed. LangChain's callback manager swallows handler exceptions and only logs a warning. Tests that assert an attribute is missing can pass even though the callback crashed. Example: test_prompt_template_omits_non_serializable_variable ([opentelemetry-instrumentation-genai-langchain] Capture prompt template name and variables on chat #730) passed only because on_llm_end raised TypeError, so the span never got its completion attributes or metrics.
• The inputs are guesses. Hand-built serialized/inputs/output arguments don't necessarily match what LangChain sends. For example, on_tool_end gets either a ToolMessage or a raw value, depending on how the tool is called.
• It doesn't match the other instrumentations. Other packages (openai, anthropic, llama-index, agno, ...) test through the library's public API, using VCR cassettes or fake models.

Proposal:

• Rewrite these tests to use real LangChain primitives: fake chat models, @tool, RunnableSequence, prompt templates, and LangGraph agents. Use cassettes where a real provider is involved.
• Make every test check that the span finished (e.g. response model, token usage), so a crashed callback can't pass silently.
• Delete the direct-handler tests once the new tests cover the same behavior. Keep a handful only where no LangChain API reaches the code path.