Add HuggingFace tokenizer.json format support (#17)

* feat: add HuggingFace tokenizer.json format loading support

Add HuggingFaceTokenizerLoader for parsing HF tokenizer.json files,
supporting both BPE (Gemma) and Unigram (Llama) SentencePiece models.
TokenizerJsonLoader now auto-detects HF format and dispatches accordingly.

Closes #10

https://claude.ai/code/session_01Qt65mqTaXaENRyhCKQYjzJ

* refactor: simplify HuggingFace loader after code review

- Extract specialContents set into _HfMetadata (avoid duplicate construction)
- Eliminate double map lookup in BPE merge score resolution
- Extract shared _buildTrainerSpec/_buildNormalizerSpec helpers
- Define special token variant constants at file level
- Remove unnecessary fromMap/_fromJsonMap indirection
- Unify test fixture builders with shared defaults

https://claude.ai/code/session_01Qt65mqTaXaENRyhCKQYjzJ

---------

Co-authored-by: Claude <noreply@anthropic.com>

Author: brody-0125

lib/dart_sentencepiece_tokenizer.dart

@@ -15,6 +15,8 @@ export 'src/sentencepiece/sentencepiece_tokenizer.dart'
         SpTruncationConfig,
         SpTruncationDirection,
         ModelType;
+export 'src/sentencepiece/serialization/huggingface_json.dart'
+    show HuggingFaceTokenizerLoader;
 export 'src/sentencepiece/serialization/tokenizer_json.dart'
     show
         SentencePieceTokenizerJson,