[Feature Request] Support Past Covariates in AutoGluon Time Series Models

## Description
I would like to request support for incorporating past covariates in the training process of time series models within AutoGluon. This feature would enhance the flexibility and predictive power of time series models by allowing them to leverage additional historical information.

- This proposal refers to the time-series module.

## Requested Enhancements:

- Enable `PatchTSTModel` and `DeepAR` to support the inclusion of past covariates during training.

- Allow fine-tuned `Chronos-Bolt` models to accept past covariates for improved forecasting capabilities.

## Motivation:

- Many real-world time series problems require contextual historical information beyond the target variable itself.

- This enhancement would enable more accurate and robust forecasting, especially for datasets with external influencing factors.

If there are any current workarounds or ongoing developments related to this, I would appreciate any insights. Thank you for considering this feature request. I appreciate the efforts of the AutoGluon team and look forward to any discussions about feasibility and potential implementation!

Start in the time-series module and inspect how PatchTSTModel, DeepAR, and fine-tuned Chronos-Bolt currently receive training data. Trace existing covariate handling, then determine the model and test entry points needed to accept past covariates and verify forecasting behavior for each requested model.

Written by the indexing model from the issue text.