Use batch_decode in WordStoppingCriteria for efficiency

[Copilot]

Replace list comprehension with individual decode() calls with batch_decode() for potential performance gains when processing batches.

Changes

• WordStoppingCriteria.__call__(): Use self.tokenizer.batch_decode(input_ids.tolist()) instead of [self.tokenizer.decode(ids) for ids in input_ids]
• Test coverage: Add test_word_stopping_criteria_utf8_tokenizer() to verify batch_decode() and decode() produce identical results with UTF8Tokenizer

# Before
texts = [self.tokenizer.decode(ids) for ids in input_ids]

# After  
texts = self.tokenizer.batch_decode(input_ids.tolist())

No behavioral changes - batch_decode() is part of the PreTrainedTokenizer interface and produces identical output to individual calls.

Original prompt

---

This section details on the original issue you should resolve

<issue_title>WordStoppingCriteria efficiency</issue_title>
<issue_description>From #2

WordStoppingCriteria can be made more efficient by using
self.tokenizer.batch_decode(input_ids.tolist()) instead of [self.tokenizer.decode(ids) for ids in input_ids]

We need to test this works with utf8-tokenizer for example, which is already a dependency - UTF8Tokenizer should have the same behavior for both options, and we should use batch_decode because it could be faster</issue_description>

<agent_instructions>implement the change, and test it</agent_instructions>

Comments on the Issue (you are @copilot in this section)

• Fixes WordStoppingCriteria efficiency #4

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.