Fix tokenizer check script: safe dataset access, default checkpoints, and tested in dry-run mode

[aijadugar]

what this PR does and summary

This PR improves the scripts/check_tokenizers.py script by:

1. Safe dataset access:

   • Handles dataset[i]["premise"] and dataset[i]["hypothesis"] safely for the English XNLI dataset.
   • Avoids errors with dictionary-like data for multi-language datasets.

2. Default checkpoints per tokenizer:

   • Provides safe default checkpoints to avoid class mismatch warnings.
   • Skips tokenizers without a defined checkpoint.

3. Dry-run tested:

   • Tested only on a small subset (5 samples) by default.
   • Avoids excessive downloading and testing during development.

4. Counters reset per tokenizer:

   • total, perfect, imperfect, wrong counters reset for each tokenizer run.
   • Accurate reporting of tokenizer comparison results.

5. Improved error handling:

   • Skips tokenizers gracefully when dependencies are missing or errors occur.

---

Example Output

[aijadugar]

hello team, PR is ready for review...

[Rocketknight1]

We're not sure we're keeping that script, unfortunately! I don't want to start merging fixes to it right now.

[aijadugar]

ok @Rocketknight1 , what I do now?