[SM90] Add optimized SM90 support.

[yuyu5333]

Summary

This PR adds an optimized SM90 build and runtime path for Hopper GPUs such as H20 and H100, while preserving the existing SM100/SM103 defaults.

SM90 uses an 8-CTA cluster for small-batch, long-sequence BF16 Top-K workloads. The cluster path is enabled from a vocabulary size of 128K, based on measurements on an 8×H20 machine. SM100 and SM103 continue to use the existing 16-CTA cluster and 512K threshold.

Changes

• Add DEEP_SELECT_CUDA_ARCHS to select 90a, 100a, and/or 103a at build time.
• Keep the default build targets unchanged: sm_100a and sm_103a.
• Add eight SM90 cluster instantiations covering:
  • int32 and int64 output indices
  • sorted_index enabled and disabled
  • return_value enabled and disabled
• Dispatch SM90 BF16 workloads to an 8-CTA cluster when:
  • batch_size <= 6
  • vocab_size >= 128K
  • topk <= 1024
• Preserve the existing 16-CTA cluster and 512K threshold for SM100/SM103.
• Fall back to nvidia-smi -L for CUDA platform detection when lspci is unavailable in a container.
• Avoid the std::format dependency in the C++ binding so the extension builds with older host standard libraries commonly found in CUDA containers.
• Document the SM90 build command:

DEEP_SELECT_CUDA_ARCHS=90a pip install -v .

Performance

Measured on NVIDIA H20 with BF16 input, batch_size 1 and 6, and topk 512 and 1024. Each result is the median of 100 cold-cache runs.

The improvement column compares the new 8-CTA cluster with the existing single-CTA path compiled for SM90.

Vocabulary size	Improvement over single CTA	Speedup over torch.topk
128K	25.2%–35.0%	3.11×–3.83×
256K	36.5%–44.6%	2.79×–3.57×
512K	52.4%–56.1%	2.66×–3.75×
1M	61.5%–62.3%	2.12×–4.11×

Short-sequence and large-batch workloads continue to use the existing normal kernel. Their measured speedup over torch.topk remains 4.88×–6.27×.

Validation

• Clean SM90 build with CUDA 13.0 and PyTorch 2.11.
• cuobjdump register-spill check passed with no spills.
• 20/20 targeted correctness cases passed.
• Coverage included:
  • normal and 8-CTA cluster paths
  • BF16 and FP32 inputs
  • int32 and int64 indices
  • sorted_index
  • optional value output
  • variable row lengths through end
  • per-row index offsets
• The final performance matrix passed Top-K value-set validation against torch.topk for every case.