[AMD] Fix DeepSeek block-FP8 loading on gfx94x

[bingxche]

Motivation

On MI300/MI325, SGLang converts checkpoint FP8 weights to the FP8 form used by the GPU. That conversion creates a new scale tensor and drops the small label that says whether the scale holds normal numbers or raw UE8M0 bytes. Later code reads the missing label and crashes before DeepSeek can start.

Seen in https://github.com/sgl-project/sglang/actions/runs/34070685413/job/101587367851

.

Modifications

• Remember whether the original scale contains raw UE8M0 bytes before conversion.
• Keep the existing conversion unchanged for normal block-FP8 scales.
• Do not treat raw UE8M0 bytes as normal numbers; let the existing requantization handle them.

This keeps non-FNUZ paths unchanged and preserves the UE8M0 support added in #36805.

Accuracy Tests

• Targeted ROCm 7.2.4 / MI300 8-GPU CI

Speed Tests and Profiling

Not applicable; this is a load-time condition fix.

Checklist

• Code is formatted.
• Documentation is not needed for this internal loader fix.
• Accuracy validation is listed above.
• The change follows the existing code style.

---

CI States

Latest PR Test (Base): ❌ Run #34242452181
Latest PR Test (Extra): ❌ Run #34242451829
Latest PR Test (AMD ROCm 10): ❌ Run #34242452279