[ROCm] Support DeepSeek V4.1 PTPC-FP8 runtime layouts

[wangjiaxin99]

This PR adds the remaining runtime support needed to run
amd/DeepSeek-V4.1-Flash-Quark-PTPC-FP8
on ROCm.

This PR is built on top of PR #57071. It relies on PR #57071 for:

• DeepSeek V4.1 Quark mixed-precision checkpoint loading.
• Quantization mapper compatibility, including routing engram.embed scale
  tensors into the engram module.
• OCP-MX/MXFP8 linear support needed by the base DeepSeek V4.1 quantized model.

This PR only adds the remaining PTPC-FP8 runtime layout support that is not part
of PR #57071.

Changes

• Support PTPC-FP8 wo_a reconstruction in the ROCm sparse MLA path.

  • Accept per-channel scale tensors.
  • Handle Quark-loaded [in, out] wo_a weights before caching the BF16
    layout used by the inverse-RoPE einsum.
  • Preserve the existing MXFP8/block-scale path.

• Support PTPC-FP8 weight-only engram.embed row-wise scales.

  • Allow 1-D FP32 row-wise scale shards in ParallelEngramEmbedding.
  • Add a Triton lookup branch for row-wise FP32 scales.
  • Preserve the existing E8M0 block-scale path.

Validation

• Docker image:

  • vllm/vllm-openai-rocm:nightly-eed1f3d0c6043bd494424a22443ee198dd56f657

• GSM8K full eval:

  • Original DeepSeek V4.1 Flash model:
    • flexible-extract: 0.9295
    • strict-match: 0.9295
  • amd/DeepSeek-V4.1-Flash-Quark-PTPC-FP8:
    • flexible-extract: 0.9295
    • strict-match: 0.9287

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @wangjiaxin99.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork