[DEV] feat: Enable GPU Marlin repack in InfiniLM AWQ/GPTQ weight processing

[qinyiqun]

Background

InfiniLM currently needs Marlin-compatible weight layouts for AWQ/GPTQ inference. CPU-side repack works but increases model loading cost and does not use the new InfiniCore GPU repack operators.

Scope

• Use InfiniCore AWQ/GPTQ Marlin repack operators when available.
• Keep fallback behavior for environments without Marlin support.
• Integrate the repacked weights into the existing quantized linear process-weight flow.

Follow-up Work

• Improve workspace/cache reuse.
• Optimize server integration.
• Reduce communication overhead for tensor parallel inference.
• Continue profiling decode performance against vLLM.