Add various difficulty evals and a UI for comparison

[andrewginns]

This PR introduces evaluation modules for mermaid diagram fixing across various levels of difficulty, adds robust multi-model benchmarking, and provides a Streamlit UI dashboard for interactive comparison. It expands the repo’s ability to evaluate, compare, and visualise agentic model performance using multiple MCP servers and tool integrations.

Problem:

• Previous evaluation functionality was limited to basic or single-case scenarios.
• No systematic way to benchmark multiple models or difficulty levels.
• Lacked a user interface for visualising evaluation results and resource usage.
• Cost calculation, error handling, and metrics aggregation were inconsistent or manual.

Solution:

• Added a new evaluation suite covering easy, medium, and hard cases for mermaid diagram correction using multiple MCP servers.
• Implemented parallel and sequential evaluation runners for multi-model benchmarking (run_multi_evals.py).
• Introduced comprehensive error/retry handling and granular metrics collection (usage, tokens, tool calls, failure reasons).
• Developed a Streamlit dashboard (merbench_ui.py) and a configurable UI backend for result visualisation, filtering, cost analysis, and leaderboard generation.
• Added cost schema/configuration and CSV result exporting for downstream analysis.
• Improved docs and in-code comments for clarity and extensibility.

Unlocks:

• Enables robust, repeatable benchmarking across LLM models and agent frameworks.
• Provides a rich dashboard for model comparison by accuracy, cost, speed, and resource usage.
• Simplifies diagnosing agent/tool failures and supports deeper analysis of model/tool interactions.
• Facilitates future extension to other agent frameworks, evaluation types, or visualisation needs.

Detailed breakdown of changes:

• Added new directory: agents_mcp_usage/multi_mcp/eval_multi_mcp/ with:
  • evals_pydantic_mcp.py: Core evaluation logic supporting multi-level difficulty, retry logic, rich metrics, and robust error handling.
  • run_multi_evals.py: Script to run evaluations across multiple models, supports both parallel and sequential execution, writes results to CSV.
  • dashboard_config.py: Configuration for the new UI dashboard, specifying metrics, grouping, cost breakdown, and plot settings.
  • merbench_ui.py: Streamlit-based user dashboard for visualising and comparing evaluation results.
  • costs.csv: Pricing table for supported models (used for cost calculation in dashboard and reports).
  • README.md: Comprehensive documentation for the new evaluation system, covering usage, error handling, output schema, and available options.
• Enhanced agent scripts (basic_mcp_use/*.py) with improved docstrings and argument clarity.
• Added/fixed extensive inline documentation and error-handling comments throughout new modules.
• All new features and files are modular, extensible, and designed to support both current and future MCP evaluation scenarios.