[Fix/Feature] Fix code for huggingface bitsandbites and auto GPTQ

[b-re-w]

Issue 1: AutoGPTQ backend remains a script and isn’t integrated as a runtime

• api/src/main/backend/gptq.py still contains installation commands and example code instead of a CoreRuntime subclass, so it cannot be used like the other backends.

Implement and register AutoGPTQ runtime

1. Remove script-style content from api/src/main/backend/gptq.py and introduce a GPTQRuntime class inheriting from CoreRuntime.

   • Load models via AutoGPTQForCausalLM.from_quantized and implement a __call__ method that provides both streaming and non-streaming generation similar to BinRuntime.

2. Register the backend at the bottom of the file with CoreRuntime.register_backend("GPTQRuntime", GPTQRuntime).

3. Update api/src/main/backend/__init__.py:

   • Add GPTQ = "GPTQRuntime" to the BackendType enum.
   • Import GPTQRuntime with from .gptq import GPTQRuntime.

4. Modify GPTQ-capable model classes (e.g., api/src/main/models/qwen3/model.py) to include BackendType.GPTQ in supported_backends and handle it in _get_runtime.

Issue 2: BinRuntime lacks consistent cache handling and flexible quantization settings

• Although BinRuntime defines a local cache path, it uses None for cache_dir in from_pretrained, causing inconsistent cache usage, and it doesn’t save the tokenizer locally.

Improve BinRuntime caching and quantization options

1. In api/src/main/backend/bin.py, pass cache_dir or self.__cache_dir to both AutoModelForCausalLM.from_pretrained and AutoTokenizer.from_pretrained, and include trust_remote_code=True.
2. Save the tokenizer to the same cache path with tokenizer.save_pretrained(save_path) whenever downloading a model.
3. Expose quantization_config as an optional constructor argument and document how to adjust BitsAndBytesConfig settings.

Issue 3: AutoGPTQ dependency missing from pyproject.toml

• pyproject.toml lists bitsandbytes but not auto-gptq, so installations fail when the AutoGPTQ backend is used.

Add AutoGPTQ dependency

1. Add "auto-gptq>=0.x" to the [project] dependencies section in pyproject.toml.
2. Update documentation (e.g., README) to mention AutoGPTQ requirements and GPU/CUDA compatibility notes.