feat: add compare_estimators_tool for agentic model selection

[rupeshca007]

Summary

Closes #369

What this adds

A new compare_estimators_tool MCP tool that enables agentic model selection directly from natural language prompts.

How it works

1. LLM agent receives: "Find best forecasting model for this dataset"
2. 2. Agent calls compare_estimators(dataset="airline", estimator_names=[...])
3. 3. Tool runs expanding-window cross-validation on each model
4. 4. Returns a ranked leaderboard with the best model identified

Example output

{
  "success": true,
  "best_model": "NaiveForecaster",
  "best_score": 0.127075,
  "leaderboard": [
    {"rank": 1, "estimator": "NaiveForecaster", "test_MeanAbsolutePercentageError": 0.127075, "status": "ok"},
    {"rank": 2, "estimator": "ThetaForecaster", "test_MeanAbsolutePercentageError": 0.129265, "status": "ok"}
  ],
  "metric_used": "test_MeanAbsolutePercentageError",
  "cv_folds": 3,
  "dataset": "airline",
  "failed": []
}

Files changed

• src/sktime_mcp/tools/compare_estimators.py - new tool (builds on evaluate_estimator_tool)
• • src/sktime_mcp/tools/__init__.py - registered compare_estimators_tool
• • src/sktime_mcp/server.py - registered tool in list_tools() and call_tool() dispatcher