Implement modular encoder/decoder class

[bkarab03]

This PR implements a set of encoder/decoder classes that provide a consistent interface for encoding and decoding text using different schemes like character or BPE encoding.

The main changes:

• Add EncoderDecoder base class with encode()/decode() methods
• Implement CharEncoderDecoder and BPEEncoderDecoder subclasses
• CharEncoderDecoder loads mappings from metadata file
• BPEEncoderDecoder uses tiktoken for tokenization
• EncoderDecoder picks correct subclass based on metadata

This allows cleanly switching between different encoding implementations by just changing the imported EncoderDecoder class. New schemes can easily be added by implementing a new subclass.

Some benefits:

• Consistent usage of encdec.encode() and encdec.decode()
• Encoder implementation details are hidden
• Easy to add new encoding types
• Model code does not change, just the EncoderDecoder class

Let me know if any changes are needed! Please take a look at the EncoderDecoder class design and see if it meets the goals of modular and extensible encodings.

Usage example:

encdec = EncoderDecoder(meta_path)
tokens = encdec.encode(text)
text = encdec.decode(tokens)```