feat: Introduce Task-Aware Execution Tools for Classification and Regression

[amruth6002]

Description

fixes #46

Addresses the limitation discussed in #46 where fit_predict was severely constrained by hardcoding a forecasting horizon parameter. This prevented LLMs from utilizing sktime's rich collection of Time Series Classification (e.g., Rocket, HIVECOTEV2) and Regression models.

This PR introduces two dedicated, task-aware execution tools that allow the MCP to smoothly interact with supervised panel data out-of-the-box.

Key Changes

1. New Tools: Added fit_predict_classification and fit_predict_regression to the existing toolset via src/sktime_mcp/tools/classify.py.
2. Runtime Execution Details:
   • Created _resolve_supervised_data() in executor.py to seamlessly route execution either through new built-in demo datasets or via custom data handles passed through load_data_source.
   • Implemented evaluation metrics: the tools will automatically compute and return accuracy (for classifiers) or RMSE/MSE (for regressors) if test labels (y_test) are present in the dataset.
3. Expanded Demo Data: Added dedicated classification variants (arrow_head, gunpoint, basic_motions, italy_power_demand) and regression variants (covid_3month, cardano_sentiment) to the dataset registries.
4. Discoverability Fix: Refactored list_available_data and the underlying executor.list_datasets() to logically group demo datasets by analytical task (forecasting, classification, regression) so LLMs can easily discover what's available contextually.

Testing

• Extended tests/test_core.py to ensure unit tests pass with the refactored categorized layout of list_datasets().
• All 15 existing tests pass.