[faresobeid]

Adds a deep-select wheel to the kernel release, built from deepseek-ai/DeepSelect @ bfa4507 (v1.0.0 + Ascend).

prime-rl wants it for the GLM-5.x / DeepSeek-V4 DSA indexer top-k (fp8_indexer.py, today torch.topk over fp32 rows up to 131k wide).

• Upstream compiles only sm_100a / sm_103a. The sm_90a gencode line is commented out upstream only to save compile time. The script turns it on, so one wheel covers H100/H200 and B200/B300. There is no sm_120.
• setup.py calls lspci before it reads DEEP_SELECT_BUILD_TARGET_PLATFORM. Build containers have neither lspci nor a GPU, so the script puts a one-line lspci stub on PATH.
• Smoke test: deep_select.topk vs torch.topk (same index sets) on sm_90/100/103, skipped elsewhere.

Validated on H200 (CUDA 13.0, torch 2.13.0+cu130):

• The script builds the wheel in ~5 min with MAX_JOBS=64 (no register spills). cuobjdump shows sm_90a, sm_100a and sm_103a cubins.
• scripts/smoke_test.py deep-select passes.
• At GLM-5.3 shapes ([32k x 131k] fp32, top-2048): torch.topk 54 ms, DeepSelect 11-12 ms (3.6-10.9 ms with per-row end). The selected values match torch.topk exactly in every row; index sets differ only where logits tie at the k boundary.
• Not validated: the aarch64 build, and running on SM100/SM103 (no Blackwell here).

Consumer PR in prime-rl: PrimeIntellect-ai/prime-rl#3958 (it needs a release of this wheel before it can pin it).