fix: include user question in chunk extraction prompts

[AmeyaBadve]

Problem

The Markdown and HTML chunk extraction prompts omit the user's question, even though the standard GenerateAnswerNode passes it as an input variable.

Consequently, the LLM processes webpage content without knowing what information the user requested, potentially returning incorrect or irrelevant results.

The K-level answer node also uses these shared templates but previously passed the user prompt through format_instructions rather than question. Adding {question} to the templates therefore requires updating the K-level node as well.

Fix

• Include USER QUESTION: {question} in both Markdown and HTML chunk extraction templates.
• Update GenerateAnswerNodeKLevel to pass the user prompt through question.
• Supply the actual JSON formatting instructions through format_instructions in the K-level node.
• Add regression tests covering both templates and K-level chunk processing.

Testing

• All three regression tests pass locally (3 passed).
• Verified successful extraction using Ollama with Qwen3:8B.
• Confirmed that the K-level chunk prompt receives both the user question and JSON formatting instructions.