[Quant] Let DeepGEMM consume prequantized activations

[WoosukKwon]

Purpose

Let block-scaled DeepGEMM linear kernels consume a compatible
QuantizedActivation produced upstream instead of quantizing the same BF16
activation again.

The kernel now advertises its dynamic 1x128 FP8 input contract, preserves the
original output shape, and trims only the logical rows from compiler-padded
packed UE8M0 scale storage at the custom-op boundary.

This is an independent reusable prerequisite extracted from #51936; it does not
include MoE, all-reduce, model, or MTP changes.

Duplicate-work check

No issue number was provided. I searched open PRs for DeepGEMM prequantized activation. The only exact match was the superseded draft #51936. #49828 adds
a standalone SiLU/multiply dynamic per-token CUDA kernel and does not add this
block-scaled linear input contract.

Tests

.venv/bin/python -m pytest tests/fusion/test_quant_activation_contract.py -q
# 5 passed

pre-commit run --files \
  tests/fusion/test_quant_activation_contract.py \
  vllm/model_executor/kernels/linear/scaled_mm/BlockScaledMMLinearKernel.py \
  vllm/model_executor/kernels/linear/scaled_mm/deep_gemm.py \
  vllm/model_executor/layers/quantization/fp8.py
# all applicable hooks passed

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.