fix: harden LLM and Cypher execution

[pntech20]

Summary

• add read-only Cypher validation before generated queries reach Neo4j
• apply a configurable LLM timeout to OpenAI/CLARIN/Gemini clients and wrap async LLM calls with asyncio.wait_for
• retry guardrail and Cypher generation chains with exponential backoff
• document LLM_TIMEOUT_SECONDS and add focused guardrail tests

Verification

• uv lock
• uv run ruff format src\llm_hardening.py src\mcp_client\client.py src\mcp_server\server.py src\mcp_server\tools\knowledge_graph\rag.py src\topwr_api\server.py tests\test_cypher_guardrails.py
• uv run ruff check src\llm_hardening.py src\mcp_client\client.py src\mcp_server\server.py src\mcp_server\tools\knowledge_graph\rag.py src\topwr_api\server.py tests\test_cypher_guardrails.py
• uv run --with pytest python -m pytest tests\test_cypher_guardrails.py -q
• git diff --check

Fixes #28