Implement eq in SentencePieceModel based on __getstate__

This allows comparing two vocabularies without loading the model. Currently, eq is implemented by comparing the md5 checksum of the loaded models, which requires the model to be loaded. It also ignores other parameters of the vocabulary such as `extra_ids`, `reverse_extra_ids`, etc.

PiperOrigin-RevId: 684142783

seqio/vocabularies.py

@@ -555,15 +555,7 @@ def _decode_tf(self, ids):
   def __eq__(self, other):
     if not isinstance(other, SentencePieceVocabulary):
       return False
-    try:
-      their_md5 = hashlib.md5(other.sp_model).hexdigest()
-    # If other has no sp_model attribute, we can't test for equality
-    except AttributeError:
-      return False
-    if self.sp_model is None:
-      return False
-    our_md5 = hashlib.md5(self.sp_model).hexdigest()
-    return our_md5 == their_md5
+    return self.__getstate__() == other.__getstate__()
 
   def __str__(self) -> str:
     return (