[cutedsl] vectorize and stage aux epilogue stores

[jansel]

Stacked PRs (oldest at bottom):

• [cutedsl] Update benchmarking code #2521
• [cutedsl] add target1 direct entry path #2520
• [cutedsl] introduce target1 clc scheduler object #2519
• [cutedsl] add tcgen05 tvm ffi launch with seed configs #2518
• [cutedsl] add guarded flat tcgen05 role coordinates #2517
• [cutedsl] cache cute launch arguments #2516
• [cutedsl] add explicit tcgen05 epilogue tile and two-cta kloop #2515
• [cutedsl] introduce pure matmul object model for tcgen05 #2514
• [cutedsl] add tcgen05 lifecycle context and kloop ownership #2513
• [cutedsl] skip static full tma fallback #2512
• [cutedsl] add staged scheduler mailbox diagnostic #2511
• ->[cutedsl] vectorize and stage aux epilogue stores #2510
• [cutedsl] retile and pack aux epilogue #2509
• [cutedsl] seed aux TMA and clc persistence candidates #2508
• [cutedsl] add aux TMA load mode #2507
• [cutedsl] add clc scheduler diagnostics and lifecycle fixes #2506
• [cutedsl] split scheduler edge epilogues #2505
• [cutedsl] tune edge K-tail epilogue placements #2504
• [cutedsl] add hybrid TMA store pipeline for edge tiles #2503
• [cutedsl] enable clustered edge TMA and seed configs #2502
• [cutedsl] support flat edge TMA stores with scalar fallback #2501
• [cutedsl] handle CuTe edge epilogue autotune #2500
• [cutedsl] replace CuTe golden codegen tests #2499
• [cutedsl] split CuTe tcgen05 config logic #2498
• [cutedsl] move CuTe state out of DeviceFunction #2497

---

[cutedsl] vectorize and stage aux epilogue stores