CUDA: fix get_rows_back for tables with more than 65535 rows (grid-y clamp + stride)

[mattjallo]

Problem. ggml_cuda_op_get_rows_back launches a grid whose y-dimension is the destination row count (ne1) with no clamp. Any table with more than 65535 rows — e.g. an embedding/vocab gradient — aborts at launch with GET_ROWS_BACK failed: invalid argument (CUDA grid y/z max is 65535). The forward get_rows kernel already clamps grid dims to UINT16_MAX; the backward one didn't.

Fix. Clamp grid.y to MIN(ne1, UINT16_MAX) and grid-stride over the destination rows in k_get_rows_back_float. Results and determinism are unchanged for tables that already fit; larger tables now compute correctly (the kernel stays the existing deterministic gather — perf is unchanged).

Test. Adds a test-backend-ops case test_get_rows_back(F32, n=1, m=70000, r=4, b=1). Without the fix it aborts on CUDA; with it, GET_ROWS_BACK passes 4/4 on an RTX 4060 (verified against the CPU reference).

[ggml-gh-bot]

Hi @mattjallo, thanks for your contribution!

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

• AI-generated content: This project does not accept PRs, descriptions or commit messages that are fully or predominantly AI-generated. If you have used AI to assist you in writing code, please make sure to disclose that explicitly.

---

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

[mattjallo]

re: flag from ggml-gh-bot, yes predominately - hand reviewed and tested