[Feature]: support AutoScheme in model free mode

[xin3he]

Feature Description

AutoScheme relies on model objects and CUDA devices, which necessitates extensive modifications to the "model-free" logic.
To ensure proper functionality, additional time is required to investigate specific implementation strategies.

Motivation and Use Case

Many model cannot reach good accuracy without mix precision, such as GLM5.1, step1.1

Alternatives Considered

No response

Definition of Done

No response

Additional Context

No response

[wenhuach21]

This is not feasible. What I mean is that we should evaluate this scenario, in which users do AutoScheme first and then model free quantization. If this is not supported, we need to fall back to legacy mode.

[xin3he]

If layer_config is already generated by AutoScheme, model_free should be able to accept it seamlessly. I'll give it a try later.