[Bug]: LGBM Results Not Reproducible

[dannycg1996]

Describe the bug

Currently there are a couple of issues around the reproducibility of the FLAML best loss, when the best estimator found was an LGBMEstimator.
These issues occur as follows:

1. The best loss returned by FLAML is not reproducible, using the underlying LGBMClassifier or LGBMEstimator model (i.e. automl.model.model). This seems to be caused by n_estimators always being set to 1, regardless of what value it should be. Please note that n_estimators always seems to be set correctly on the FLAMLised LGBMEstimator - this issue exists exclusively on the underlying models.
2. With certain configurations, the FLAML best loss can't be reproduced even when using the FLAMLised LGBMEstimator. This seems to only be the case when a time budget is set, so the issue is likely caused by the callbacks, similar to the issue found with CatBoostEstimators here

Steps to reproduce

In terms of showcasing these issues in a reproducible manner:

1. Can be replicated by uncommenting and then running coverage run -m pytest -k test_reproducibility_of_underlying_regression_models[lgbm] (or the classification equivalent.
2. A unit test will be added for this, which will initially fail, until the issue is fixed.

I'll open a PR for this shortly - hope that's okay

Model Used

LGBMEstimators

Expected Behavior

Taking the FLAMLised LGBMEstimator model, or the underlying model it wraps (LGBMClassifier/LGBMRegressor) and training and testing it on the same folds, should return the same best loss as FLAML.

Screenshots and logs

No response

Additional Information

No response