[LLM:Bugfix] Pass calibration input IDs to AWQ position IDs

[ZedingZhang]

Description

• Forward the existing calibration token IDs to get_position_ids during AWQ initialization.
• This lets multimodal visual position-ID logic inspect the calibration input instead of receiving None, fixing the initial None.flatten crash reported in AWQ quantizer crashes on multimodal models: get_position_ids() called without input_ids #4804.
• Keep the change narrowly scoped; the separate hybrid linear-attention limitation described in the issue is not addressed here.

Fixes #4804

Module

LLM

Type

• Feature
• Bugfix
• Perf
• Refact
• Style
• Doc
• Test
• Chore

Validation

• Ran a focused Python regression harness with a multimodal-style get_position_ids implementation. It reproduced the original None.flatten failure before the change, then verified that the exact calibration input IDs are forwarded and AWQ initialization completes after the change.
• python -m py_compile transformers/llm/export/utils/awq_quantizer.py
• git diff --check

A full Qwen2.5-VL AWQ export was not run locally; this PR only claims the initial missing-input fix above.

Checklist

• Commit message follows [Module:Type] Description format
• Code compiles without errors
• Tested on relevant platform(s)
• No unrelated format or style changes included