[Quant] Fuse all-reduce RMSNorm with packed FP8

[WoosukKwon]

Purpose

Expose FlashInfer's all-reduce + residual + RMSNorm + packed per-group FP8
epilogue to eager model paths.

The helper returns both the normalized BF16 tensor and a compatible
QuantizedActivation, allowing an upstream model integration to skip a
separate DeepGEMM quantization kernel. It falls back to the existing
all-reduce/RMSNorm path when the pattern, topology, scale format, or consumer
contract is unsupported.

This is an independent primitive extracted from #51936. It does not change a
specific model and does not depend on #51939 or #51941; the test uses the public
consumer contract directly.

Duplicate-work check

No issue number was provided. I searched open PRs for FlashInfer allreduce RMSNorm packed FP8 and allreduce RMSNorm FP8 quant. #42230 implements a
separate SymmetricMemory Triton collective, #45364 covers Llama RMSNorm-to-quant
without all-reduce, and #45770 fixes mixed norm-weight dtypes. None expose this
FlashInfer packed UE8M0 epilogue.

Tests

.venv/bin/python -m pytest \
  tests/kernels/core/test_fused_allreduce_gemma_rms_norm.py::test_fused_allreduce_rms_norm_group_fp8 -q
# 4 passed

git diff --cached --name-only -z | xargs -0 pre-commit run --files
# all applicable hooks passed, including ruff and mypy

AI assistance disclosure

This change was developed with OpenAI Codex assistance. This is a draft PR;
the human submitter must review every changed line and confirm they understand
and can defend the change before marking it ready.

[mergify]

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

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