Redefine agentic workflow

[saranshkr]

Summary

This PR restores the missing multi-agent analytics workflow for uploaded structured data.

The backend now uses a bounded flow:

1. Planner creates the full plan
2. Query writer generates one SQL query per step
3. Executor runs steps sequentially and stores successful outputs
4. Analyzer runs once at the end to return the final answer or request one replan
5. If limits are hit, the workflow returns a clear best-effort answer

What Changed

• Restored the missing app/agent runtime
• Replaced the old compiled-plan flow with a planner -> query writer -> executor -> analyzer workflow
• Grounded planning and execution on compact uploaded-data schema/context summaries
• Added bounded retry, bounded replan, and best-effort fallback behavior
• Kept the existing API/UI response shape stable
• Removed the old deterministic grounding path and updated tests/docs accordingly

Validation

./.venv/bin/python -m pytest tests -q