DOC: Add visible warning about limitations of oversampling methods

[omlende0410]

Fixes #1101

Adds a warning note at the beginning of the over-sampling user guide.

The note highlights some important considerations when using over-sampling
methods such as SMOTE and ADASYN:

• they may have limited impact on ranking-based metrics (e.g. ROC-AUC),
• they can affect probability calibration,
• and they are not always necessary depending on the learning objective.

The warning also points users to alternative approaches such as
class_weight and sample_weight for handling class imbalance.

This change aims to make new users aware of these trade-offs and help them
choose an appropriate strategy for their specific use case.