[DSv4.1] Move the low-ratio index top-k into dsv4/low_ratio_indexer

[DarkSharpness]

(auto generated by Claude)

What

Move the DSv4.1 ratio-1/2 (low-ratio) index top-k out of deepseek_v4_backend.py into layers/attention/dsv4/low_ratio_indexer/. The attention backend only dispatches dense / candidate source / candidate consumer layers and keeps an opaque CandidateMetadata on the forward metadata.

• DenseIndexer: the plain top-k (DeepGEMM on SM100, the portable path with the fp4 decode logits kernel elsewhere) and the prefill-graph path.
• CandidateIndexer: publish_* / consume_* for prefill and decode, backed by DeepGEMM (sparse table), DeepGEMM CP (block ids) and torch (masks).
• Shared scoring helpers in deep_gemm_utils.py and torch_utils.py.

No behavior change intended; the chunked eager decode from #40637 is kept.

Validation

GB300, DeepSeek-V4.1-Flash, TP4/EP4, main vs this PR:

Config	GSM8K main	GSM8K PR
DSpark low-latency	0.875	0.875
High-throughput + --enable-decoder-swa-bounded-replay	0.900	0.910
Same + SGLANG_DSV41_TORCH_PREFILL_INDEXER=1	0.895	0.900

• The last two rows use --chunked-prefill-size 2048 --mem-fraction-static 0.906 on both sides to avoid a pre-existing flashmla_sched_meta shared-memory overflow on main.
• A 35k-token request (beyond the candidate window) succeeds in every config.
• bs=1 DSpark verify with a fixed accept length of 3: TPOT 1.948 -> 1.940 ms at 1k input, 1.970 -> 1.963 ms at 32k input; every indexer kernel is within 1 us/step of main in the GPU profile.
• The portable decode/prefill path and the CP block publish/consume are element-wise identical to main in standalone comparisons (RTX A5000).

TODO: port the unit tests of the removed modules to the new package.

Follow-ups (not in this PR)

• Move the index projections back into the layer. PrefillInputs / DecodeInputs still carry the DeepseekV41Indexer module and the backends call queries() / head_weights() themselves, while CapturedPrefillInputs already receives projected q / weights; the backends should take tensors and metadata only.
• Dispatch inside the ops. The DeepGEMM / torch choice (use_deep_gemm_* in make_full_topk_indexer / make_candidate_indexer and the _deep_gemm_* / _torch_* method pairs) belongs in kernels/ops/attention/dsv4 as BaseFusedOp backends once the JIT and torch variants share one contract.

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ✅ Run #36213835584
Latest PR Test (Extra): 🚫 Run #36214090009
Latest PR Test (AMD ROCm 10): ⏳ Run #36213835540

[DarkSharpness]

/tag-and-rerun-ci

[BBuf]

Could you add latency and peak-memory numbers for a batched long-context eager run that spans multiple chunks? That would also check that the logits buffers are freed between chunks.