FIX: validate horizon parameter in fit_predict, predict, and fit_predict_async tools

[Sandrakimiring]

Reference Issues/PRs

Fixes #354

What does this implement/fix? Explain your changes.

Three tools accepted horizon: int but passed it through with no validation:

• fit_predict_tool — passed bad horizon to executor silently
• predict_tool — range(1, horizon+1) produced fh=[] when horizon≤0
• fit_predict_async_tool — created a background job with a bad horizon

Added a guard as the first check in each function — before any executor,
asyncio, or job-manager code is reached. Values less than 1 now return:
{"success": False, "error": "horizon must be a positive integer."}

Does your contribution introduce a new dependency? If yes, which one?

No new dependencies.

What should a reviewer concentrate their feedback on?

The 3 guards in src/sktime_mcp/tools/fit_predict.py and the 9 new
tests in TestHorizonValidation in tests/test_core.py.

PR checklist

• I've added unit tests and made sure they pass locally.