Add JSON-based serialization logic and deepcopy implementation for Sentence object

[alanakbik]

This pull request introduces a comprehensive solution for serializing and deep-copying Flair's Sentence objects, addressing a long-standing issue where standard methods like copy.deepcopy() and pickle would fail.

Closes #3245

This PR implements two key mechanisms to solve this problem:

1. JSON Serialization (to_dict / from_dict):

   • A new sentence.to_dict() method serializes the entire Sentence object—including its text, tokenizer, and all annotations (spans, relations, and labels at all levels)—into a clean, JSON-compatible dictionary.
   • A corresponding class method, Sentence.from_dict(), fully reconstructs the Sentence object from this dictionary, correctly restoring all annotations and the original tokenizer.
   • The core logic is encapsulated in two powerful private helpers, _capture_annotations() and _reapply_annotations(), which use character offsets to robustly map annotations between different tokenizations. This also allowed for a significant simplification of the existing retokenize() logic.

2. Custom Deep Copy (__deepcopy__):

   • A custom __deepcopy__ method has been implemented for the Sentence class.
   • This method intelligently navigates the object graph, manually recreating tokens, spans, and relations while respecting the caching mechanism of Span and Relation.
   • This makes Sentence objects fully compatible with Python's copy.deepcopy(), resolving the original TypeError.

Key Changes

• New Public Methods: Added Sentence.to_dict(), Sentence.from_dict(), and Sentence.__deepcopy__().
• Refactoring:
  • Refactored the Sentence constructor (__init__) to correctly handle initialization from raw text, a list of strings, and a list of pre-made Token objects, which was critical for the deserialization logic.
  • Simplified the retokenize() implementation by reusing the new annotation helper methods.
  • Made the _add_token() method more robust in calculating token start positions.
• Testing:
  • Created a new test file, tests/test_sentence_serialization.py, with comprehensive unit tests.
  • These tests verify that both deep-copying and the JSON serialization cycle perfectly preserve complex annotations (including multi-word spans and relations) and the sentence's tokenizer configuration.
• Bug Fixes: Addressed several mypy type-checking errors that arose during the refactoring to ensure the new code is robust and maintainable.