CUDA: add GGML_OP_LIGHTNING_INDEXER kernel (F16/F32 K)

[Dorijan10]

What

Adds the CUDA backend for GGML_OP_LIGHTNING_INDEXER, introduced in #24231
with a CPU implementation only. On GPU-resident DeepSeek-V3.2 / GLM-5.2 models
the fused lightning-indexer path currently has no CUDA kernel.

Implementation

Naive one-thread-per-KV-position kernel: each thread computes
sum_h relu(dot(q_h, k)) * w_h + mask over the indexer heads, matching the
CPU reference in ggml_compute_forward_lightning_indexer. Supports F16 and
F32 K; quantized K (Q8_0 etc.) to follow in a subsequent change.

Validation

Passes the existing test-backend-ops LIGHTNING_INDEXER case at the test's
1e-6 NMSE tolerance across all tested shapes (varying heads, KV, batch,
streams, mask broadcast), on an RTX 5090 (sm_120):

test-backend-ops test -o LIGHTNING_INDEXER -b CUDA0
... all f16/f32 cases: OK

Quantized K types report not supported via supports_op and are skipped.

Notes

This is a correctness-first implementation; optimization (shared-memory
q/w staging, coalesced K reads) can follow. Happy to iterate.

[ggml-gh-bot]

Hi @Dorijan10, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[abc-nix]

You know that #25545 exists, right?

[Dorijan10]

@abc-nix

Thanks I noticed it now, I knew about #25407 and #24231 (hence I got the idea from), but not this one.

I see it is more thorough. I have 8 RTX 5090 (sm_120), let me know if you guys want me to do any validation on these as well.