Feature/configurable system prompt master

[Dhara124]

Description

This PR introduces the Configurable System Prompt feature for LLM models and fixes several related interaction bugs. The changes have been strictly cherry-picked to target master and exclude unrelated dev history.

Key Changes

• Project Metadata Configuration: System prompts can now be explicitly configured via project.metadata_json.
• Per-Model System Prompts: Supports providing a dictionary for system_prompt to allow per-model overrides.
• Empty Models Set Fallback: Modified get_all_llm_output to fallback to project_metadata.get("models_set") when task["models_to_run"] is empty.
• Graceful Error Handling: Handled Llama2 API errors gracefully inside get_all_model_output, bubbling up empty interactions instead of throwing uncaught exceptions.
• Robust JSON Parsing: Ensured json.loads safely handles empty strings ("") in task interaction endpoints to prevent runtime errors.
• Punctuation Fixes: Updated DEFAULT_SYSTEM_PROMPT to use grammatically correct punctuation.

Issue Linked

Fixes issues regarding hardcoded system prompts and application crashes when dealing with missing model responses or empty interactions arrays.

Verification

• Manual verification performed to ensure conflicts were correctly resolved and feature logic remained fully intact.
• Verified that LLM outputs gracefully parse empty models arrays and fall back properly.
• All code syntax correctly checks out.