perf(nvidia): add opt-in Hopper long-context routing

[CherryLemon]

Summary

• add an architecture-specific dispatch profile for NVIDIA Hopper
• route full-attention to the CUDA vendor backend so FlashAttention-3 is selected
• blacklist FlagGems mm and mm_out so dense projections use native CUDA GEMM
• keep the optimization disabled by default; enable it with VLLM_FL_HOPPER_LONG_CONTEXT_OPT=1 for direct A/B testing
• preserve the existing NVIDIA configuration on non-Hopper devices and when capability probing is unavailable

Motivation

Nsight Systems analysis of a 16k-input serving workload showed two independent Hopper bottlenecks:

• Triton unified-attention full prefill was about 11x slower than FA3 for the profiled head shape and accounted for 76.2% of the GPU-time gap.
• FlagGems Hopper host-TMA aten::mm.out was slower than native dense GEMM for the projection shapes.

The measured output throughput moved from 535.340 tok/s to 691.895 tok/s with the attention route alone, and to 785.631 tok/s with both routes enabled, reaching 96.09% of the native baseline.

Usage

# baseline behavior, also the default
unset VLLM_FL_HOPPER_LONG_CONTEXT_OPT

# Hopper optimized route
export VLLM_FL_HOPPER_LONG_CONTEXT_OPT=1

Validation

• 44 dispatch/config/FlagGems unit tests passed in the vLLM 0.24.0 stack.
• On an H100 process, flag off selected nvidia.yaml, FlagOS attention, and no mm/mm_out blacklist.
• On the same H100 process, flag on selected nvidia_hopper.yaml, vendor attention, and the mm/mm_out blacklist.
• Ruff and git diff --check passed.