[MoE] Accept prequantized DeepGEMM inputs

[WoosukKwon]

Purpose

Allow modular FP8 MoE execution to consume a matching packed DeepGEMM
activation produced upstream, avoiding duplicate input quantization.

This PR adds an explicit prequantized-input capability to modular prepare/finalize
and expert interfaces, transports the data and scale through the opaque MoE op,
and teaches DeepGEMM's expert scatter to accept already-packed UE8M0 scales.
Unsupported dispatch modes retain the existing path.

This is independent of the block-linear consumer in #51939 and does not include
model-level all-reduce or MTP changes.

Duplicate-work check

No issue number was provided. I searched open PRs for MoE prequantized input
and DeepGEMM prequantized activation; the only exact prior implementation was
the superseded draft #51936. #49636 concerns a FlashInfer MoE-EP backend and
does not add this modular DeepGEMM input contract.

Tests

.venv/bin/python -m pytest \
  tests/kernels/moe/test_deepgemm.py \
  tests/kernels/moe/test_silu_mul_fp8_quant_deep_gemm.py -q
# 32 passed, 3 skipped

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