fix: normalize multimodal calibration inputs

[ZX-ModelCloud]

Summary

• normalize raw text calibration samples into chat conversations before invoking multimodal processor chat templates
• preserve existing structured multimodal conversations across supported VL and Omni model definitions
• place captured input_ids on the materialized input embedding device to prevent CPU/CUDA mismatches during text-only calibration
• add shared and Intern-S2 regression coverage

Validation

• pytest tests/test_prepare_dataset.py -q
• pytest tests/test_stage_inputs_capture.py tests/models/test_intern_s2_preview.py -q -k "not TestInternS2Preview"
• targeted calibration-device tests
• CUDA device-alignment smoke test
• Ruff and compile checks

Fixes #3048