feat: implement AgenticForecaster and fix evaluate tool logic

[Omc12]

Important

This PR is a submission for the sktime Agentic Track stipend.
It implements a complete "Agentic Forecaster" application and addresses critical tool improvements discovered during development.

Overview

This PR contributes to the sktime Agentic Track by implementing a functional AgenticForecaster and fixing a critical bug in the evaluation tool discovered during setup.

Changes

• AgenticForecaster: Implemented a new estimator in agentic_forecaster/ that uses natural language prompts to select sktime models via the MCP registry.
• Evaluation Fix: Resolved a bug in src/sktime_mcp/tools/evaluate.py where cv_folds was being ignored, causing incorrect fold counts in cross-validation.
• New Example: Added agentic_forecaster/main.py which demonstrates the agentic workflow on the Corporación Favorita Grocery Sales dataset.
• Git Maintenance: Updated .gitignore to exclude large data files and virtual environments to maintain repository health.

Verification

• Ran pytest tests/test_evaluate.py (Confirmed fix).
• Successfully executed main.py to produce a 30-day forecast driven by a natural language prompt.

Fixes #335
Closes to #336

[Omc12]

Hi maintainers, I noticed that PR #337 also addresses the docstring truncation bug (#335).

I wanted to clarify that my PR #338 includes this fix as part of a larger contribution for the sktime Agentic Track stipend. In addition to the registry fix, this PR implements:

• A new AgenticForecaster application.
• Full exogenous (X) handle support for the high-level fit_predict tools.
• A retail forecasting demo (main.py).

If #337 is merged first, I am happy to rebase and remove my overlapping changes, but I would appreciate a review of the Agentic features!