GH-3614: Fix identify_dynamic_embeddings for composite DataPoints

[alanakbik]

Closes #3614

This PR fixes an issue where identify_dynamic_embeddings did not correctly detect dynamic embeddings (those with requires_grad=True) within composite DataPoint types like DataPair or Sentence with token embeddings.

The logic has been refactored by:

1. Adding internal helper methods (_get_dynamic_embedding_names and _get_all_embedding_names) to the DataPoint base class with default implementations.
2. Overriding these methods in composite classes (Sentence, Span, DataPair, DataTriple) to recursively check their constituent parts.
3. Simplifying the identify_dynamic_embeddings function in training_utils.py to use these helpers.

This ensures all relevant dynamic embeddings are correctly identified across different DataPoint structures. Unit tests have been added in tests/test_training_utils.py to cover various scenarios and verify the fix.