[Core][SM70] Preserve DFlash2 decode across NVFP4 prefill

[yangzhuxinyzx]

Purpose

• keep the accepted SM70 NVFP4 QPN2 decode graph opaque while adding the QPN2-packed large-M prefill route
• replace the Python M-dependent branch and persistent 85 MiB workspace with one C++ dispatcher and an operator-local ephemeral workspace
• source-default the quality-audited 18 ms NVFP4 DFlash2 TP4/B1 practical path, preserving every explicit override and keeping candidate-order disabled

Test Plan

• focused CPU contract and dispatch tests
• exact V100 normal/gated M8/M16 operator comparisons
• restart the four-V100 256K practical API with QPN2/prefill/DFlash2 fastpath booleans removed from the launch file
• measure three 512-token single-request decode runs, cold 32K prefill, and post-prefill decode
• smoke tool calls and strict JSON schema output

Test Result

• 7 passed focused pytest; ruff, compileall, and git diff --check pass
• V100 operator comparison: max_abs=0 for normal/gated and explicit/ephemeral workspace cases
• CUDA Graph pool: 0.13 GiB
• complete DFlash2 round: 17.4035 / 17.3569 / 17.3697 ms, mean 17.3767 ms
• cold 32K pure prefill: 4039.49 tok/s
• post-32K complete round: 17.3640 ms
• tool calling: 3/3; strict JSON schema: 1/1

Detailed evidence is recorded in docs/design/sm70_dflash2_nvfp4_prefill_promotion.md.