fix(agents): parallelize LLM calls and clean handle_chat (#78, #80)

[zohaib-7035]

Summary

This PR addresses two maintenance issues regarding LLM call latency and chat logic structure inside backend/agents.py. This is an isolated refactoring, not altering the external API structure.

Changes Made

• Performance (Fixes perf: Parallelize LLM calls in extract_keywords_and_rewrite to reduce latency #78): Refactored extract_keywords_and_rewrite to parallelize the Gemini LLM calls using asyncio.gather. detect_intents and rewrite_with_history are now executed concurrently, followed by the keyword extraction running concurrently alongside the second intent detection. This mathematically optimizes latency by cutting out sequential blocking.
• Code Quality (Fixes bug: Duplicate more-query condition in handle_chat #80): Cleaned up the confusing and somewhat redundant branch check inside handle_chat. By swapping the evaluation order to check query.strip().lower() in {...} before evaluating more_count, the logic becomes safer and no longer creates a dead-path confusion with _is_more_query.

Verification

• py_compile: backend/agents.py ✔
• No missing dependencies.
• Native async code applied safely.