[Kernel][SM70] Add opt-in DeepSeek V4 FP13 GEMV

[yangzhuxinyzx]

Purpose

Add an opt-in packed-FP13 single-token GEMV for the fixed DeepSeek V4 SM70 router/compressor/indexer projections. Packing happens once after loading and retains the original FP16 GEMV fallback.

Test Plan

• CPU bit-level pack/unpack audit
• SM70 offline Triton compile and SASS resource audit
• Exclusive-V100 CUDA Graph correctness and four-stream overlap screen
• Matched PP2 x TP4 endpoint only if the overlap screen wins
• Pinned aggregate quality gates before acceptance

Test Result

• ruff check: passed
• py_compile: passed
• targeted CPU pytest: 3 passed
• offline SM70 kernel: 29 registers, 227 static instructions, 17 global loads, 32 FFMA, 8 shuffles, 2 barriers
• real-weight CPU audit: 99.87%--99.98% exact FP16 values; relative weight L2 1.18e-8--3.46e-7
• GPU overlap/endpoint gates: pending shared-GPU availability

Draft: do not merge until the queued exclusive-GPU and quality gates complete.

[yangzhuxinyzx]

Superseded by merged PR #325 (8700861e). The replacement replayed only this PR’s final FP13 kernel onto latest main, preserved the GPU-screened kernel byte-for-byte, added same-CUDA-device and tensor-value safety gates, audited all 188 eligible checkpoint tensors, quantified PP2 memory overhead, defaulted the measured path with explicit =0 rollback, and passed local plus remote static gates. The eight stale stacked commits were intentionally not replayed.

[yangzhuxinyzx]

公开开发已恢复：本 PR 已取消归档、解锁并重新打开。为避免把旧基线上的问题直接带入最新 main，现统一作为 Draft 审计项；恢复可见性不代表已经批准合并。 该组涉及 SM70/V100 性能或数值路径，需先与当前主线及 #344–#351 去重，再用同条件的聚焦质量和性能证据决定修复后合并。

[yangzhuxinyzx]

审计结论：默认开启且带保护的 FP13 路径已由后续安全实现吸收；该分支已通过修复 PR #372 同步到最新 main，最新完整 pre-commit 通过，且源码树与 main 完全一致。按零差异历史归档合并。