Fix Pascal flashprefill multi-arch link

[davide221]

Summary

• keep Pascal scalar FlashPrefill kernels visible across CUDA arch passes so default multi-arch builds do not emit host launchers with missing device symbols
• use full-warp sync vote/shuffle intrinsics and unique Pascal kernel names to avoid symbol collisions with Volta F16 kernels
• switch Qwen3.6 DFlash default docs/scripts to the uploaded Q4_K_M draft GGUF

Verification

• RunPod A40, CUDA 12.4: cmake --build server/build --target dflash_server -j8 reached and passed final dflash_server link
• local: git diff --check
• local: bash -n harness/clients/common.sh harness/benchmarks/run_lucebox_vs_llamacpp.sh
• local: python3 -m py_compile server/scripts/bench_agent.py server/scripts/bench_he.py server/scripts/bench_llm.py

[cubic-dev-ai]

No issues found across 13 files

_{Re-trigger cubic}