Fix export_dcp.py bugs

[ccchow]

This pull request introduces improvements to the model export process and checkpoint handling. The key updates include simplifying the get_model function call, ensuring shard count is at least 1, and enabling the use of cloudpickle for saving and loading checkpoint files to enhance serialization flexibility.

Model Export Improvements:

• Simplified the get_model function call by passing the config object and tokenizer length directly, reducing the number of parameters. (scripts/export_dcp.py, scripts/export_dcp.pyL138-R139)
• Ensured that the number of shards (num_shards) is at least 1 to avoid potential zero-shard errors. (scripts/export_dcp.py, scripts/export_dcp.pyL165-R162)

Checkpoint Handling Enhancements:

• Introduced cloudpickle as the pickle_module for torch.save and torch.load operations, improving serialization support for complex objects. (src/zeroband/checkpoint.py, [1] [2] [3] [4] [5]
• Fixed a logging attribute by replacing self.logger with self._logger in the delete_topk call to align with the correct attribute name. (src/zeroband/checkpoint.py, src/zeroband/checkpoint.pyL357-R358)