[BUG] SM120 dense HSTU smem overflow

[Clebrate]

Describe the bug

• On SM120 (RTX PRO 6000), dense HSTU varlen fwd (head_dim=256, bf16, no paged KV) fails with CUDA error: invalid argument at hstu_fwd.h:830.

• Actual tile is Arch80 {128,96,8} (160KB smem). SM120 per-block opt-in is 99KB.

• Default --disable_kvcache with CUDA graphs on still works: capture/replay is paged (page_size=32 → {128,32,8}, 96KB).

Steps/Code to reproduce bug
Set use_cudagraph=False in inference_benchmark.py, then: cd examples/hstu && python3 ./inference/benchmark/inference_benchmark.py --disable_kvcache

Expected behavior

• SM120 dense should use {64,64,4}.

• Map arch 120 → 89 in ARCH_SWITCH.

Environment details (please complete the following information):

• RTX PRO 6000 Blackwell
• Docker: devel_latest