Fix shared callback state in parallel OptunaSearchCV with LightGBM

[Quant-Quasar]

Motivation

When using OptunaSearchCV with n_jobs > 1, users may pass stateful callback objects
(e.g., LightGBM early stopping callbacks) via fit_params.

Because fit_params is shared across parallel workers, callback objects inside it can
be shared by reference, leading to unintended cross-trial interference.
In practice, this can cause multiple parallel trials to observe and mutate the same
callback state (such as best_iteration), resulting in incorrect or identical early
stopping behavior across trials.

This issue was reported in #240 and can silently affect optimization results when running
parallel cross-validation with callbacks.

Description of the changes

This PR ensures proper isolation of callback state across parallel trials by:

• Creating a shallow copy of fit_params at objective execution time.
• Deep-copying only the callbacks entry (if present) to provide each trial with an
  independent callback instance.
• Leaving all other entries in fit_params unchanged to avoid unnecessary duplication
  of large or expensive objects.

The change is localized to the objective execution path and does not affect existing
behavior for single-threaded runs or workflows that do not use callbacks.

This prevents shared mutable state between parallel trials while keeping memory and
performance overhead minimal.

[not522]

Thank you for your PR! Since we're approaching the year-end and New Year holidays, I'll review the details after New Year's Day.

[Quant-Quasar]

Thank you for the update! I completely understand.
Please take your time, and I’m happy to follow up after New Year’s.

[github-actions]

This pull request has not seen any recent activity.