[cutedsl] Search the KV tile width and seed the sm_103 dense shapes at 160

[yushangdi]

Stacked PRs:

• [noland] Hillclimb plan, prompt, and per-step records for CuTe attention on GB300 #3579
• [cutedsl] Retune the sm_103 flash seeds onto the one-CTA and resident lowerings #3578
• [cutedsl] Gate the causal resident softmax on its schedule, and close two KV-width holes #3577
• ->[cutedsl] Search the KV tile width and seed the sm_103 dense shapes at 160 #3576
• [cutedsl] Make the flash KV tile width a real dimension #3575
• [cutedsl] Run the dense resident softmax on the one-CTA FA4 pipeline #3574
• [cutedsl] Gate the dense flash softmax lowering on its schedule, not a seed match #3573
• Warm the GPU to its steady clock before the autotuner's final verification #3572
• Add an fa4-tuned attention baseline to the backend comparison #3571

---

[cutedsl] Search the KV tile width and seed the sm_103 dense shapes at 160

Add 160 to the searchable KV tile widths and promote it in the sm_103 dense
policy, together with the deeper staging (kv_stage 6) the wider tile leaves
room for. The choice set is deliberately constant across sequence lengths:
gating it on divisibility would make the search surface length-dependent,
and the masked trailing tile makes the width legal at any length anyway.

Dense 256K also moves off packed exp2 onto the resident softmax, which is
what the search now finds for it once the seed cliff is gone.