Unified and configurable dataset specfication

[zqhuang211]

Currently, we have two methods for specifying datasets:

• Hardcoded Classes: Each dataset has its own class inherited from the VoiceDataset class. Adding new datasets or modifying existing ones requires changing the source code.
• GenericVoiceDataset via YAML config: This approach is flexible and allows for configuration through YAML files, but it cannot yet replace all hardcoded dataset classes due to some datasets requiring special handling.

Ideally, we want to update the GenericVoiceDataset class to support a unified and configurable dataset specification through YAML configuration for all training, validation, and evaluation datasets.