Fix loading tokenizers from s3

[meilame-tayebjee]

This pull request introduces improvements to the tokenizer loading and saving logic, particularly around loading from S3 and configuration-based instantiation. The changes add a unified load_from_s3 interface to the tokenizer base class, refactor the S3 loading implementation for HuggingFace and NGram tokenizers, and introduce a build_from_config method for NGram tokenizers to streamline instantiation from configuration files.

Tokenizer loading improvements:

• Added an abstract load_from_s3 class method to BaseTokenizer to enforce a consistent S3 loading interface across all tokenizer subclasses.
• Refactored the HuggingFaceTokenizer.load_from_s3 method to simplify error messages and ensure the correct tokenizer object is instantiated from S3.
• Implemented load_from_s3 for NGramTokenizer, allowing loading from S3 using the provided filesystem and configuration, and refactored the loading logic for clarity.

Configuration-based instantiation:

• Added a build_from_config class method to NGramTokenizer to streamline tokenizer creation from a configuration dictionary, and refactored existing loading methods to use this new method.
• Removed redundant print statement from the NGram tokenizer's loading logic for cleaner output.