[ENH] Add exog_handle support to fit_predict for multivariate forecasting

[SHIVANSH-ux-ys]

Reference Issues/PRs

Fixes #336

What does this implement/fix? Explain your changes.

The fit_predict tool previously accepted only a single data_handle for the target series y. This made multivariate forecasting (where exogenous covariates like promotions or holidays are needed) impossible without manually merging datasets into a single handle.

This PR adds an optional exog_handle parameter to fit_predict and fit_predict_async (both the tool wrappers and the core executor). When provided, the covariates are resolved from _data_handles and passed as X to both estimator.fit() and estimator.predict().

Changes:

• executor.py: fit_predict and fit_predict_async now accept exog_handle and resolve it to X before fitting.
• tools/fit_predict.py: Updated fit_predict_tool and fit_predict_async_tool signatures and docstrings.
• server.py: Updated the JSON schema for both fit_predict and fit_predict_async tools to expose exog_handle to LLM agents.

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

No.

What should a reviewer concentrate their feedback on?

The exog_handle resolution logic in executor.py and whether the fallback (X = exog_info.get("y") when no explicit X column is set) is the correct behavior.

PR checklist

• I have read the CONTRIBUTING guide.
• I have checked that my PR does not duplicate an existing PR.
• I have checked that there is an existing issue for this PR, if applicable.
• My code follows the project's coding standards.

[SHIVANSH-ux-ys]

Closing this — I just noticed Omc12 had already committed a fix for this issue 3 days ago via PR #338. Apologies for the overlap, I should have checked the issue timeline more carefully before submitting.