[Feature Request] Add support for dynamic encoding registration to extend harmony format to other LLM models

[madroidmaq]

🎯 Problem Description

• The current library only supports the hard-coded HarmonyGptOss format.
• Request to add dynamic encoding registration support for extending compatibility with other LLM models.
• Maintain a unified API and streaming parsing architecture.

💡 Motivation

• Increase the adoption rate and versatility of the library.
• Reuse an excellent streaming parser architecture.
• Standardize parsing logic across different LLM providers.
• Ensure backward compatibility.

🔧 Proposed Solution

• Configuration-driven system: Define encoding parameters through YAML/JSON configuration files.
• Plugin registration system: Ability to register new formats at runtime.
• Expected API: load_harmony_encoding("llama3")

📈 Benefits

• Backward compatibility
• Unified API
• Extensibility
• Friendly to community contributions

🛠️ Technical Considerations

• The existing state machine architecture is applicable to other LLM formats.
• The three-channel system can accommodate different models.
• Token-level streaming processing is universally suitable.