[Feature]: refine lm-head quantization

### Feature Description

LM heads are getting larger and larger, while quantizing the LM head can be very expensive in terms of memory and runtime.

Calibration: For some popular models, use explicit model architectures to fetch the lm-head inputs directly from the last transformer block, instead of running calibration through the entire model.

Chunk-wise processing: Use NVIDIA's, Unsloth's, or Triton-based implementations to process large layers chunk by chunk during the forward/backward pass, reducing peak memory usage.

### Motivation and Use Case

~

### Alternatives Considered

~

### Definition of Done

_No response_

### Additional Context

_No response_

Start by mapping the existing calibration and LM-head quantization paths; the issue does not name files, tests, or entry points. Compare the proposed explicit model-architecture calibration with NVIDIA, Unsloth, and Triton chunk-wise implementations, and clarify a Definition of Done covering reduced peak memory and runtime.

Written by the indexing model from the issue text.