Feat/langfuse integration + langfuse v3

[fidok15]

Langfuse was wired up but the data was useless — one global CallbackHandler meant every request wrote into the same trace, concurrent requests overwrote each other's handler on the RAG object, and the final-answer LLM call wasn't traced at all.

What changed:

• trace_id is generated per request in the ToPWR API and passed down to the MCP server
• handler is created per request and travels through LangGraph state instead of self.handler, so parallel requests can't mix traces
• moved handler setup to the Langfuse v3 API (trace_context — the old constructor kwargs crash on v3)
• final answer call is traced under the same trace_id, so one user request = one trace: Guardrails → Generate Cypher + Final Answer
• real session_id is propagated end to end (was: trace_id reused as session), so conversations group correctly in the Sessions view
• both servers flush Langfuse on shutdown (lifespan / atexit)