cuda : add F16 kernel support for CONV_2D_DW

[crion99]

Overview

This PR adds F16 kernel support for GGML_OP_CONV_2D_DW in CUDA.

Originally, CUDA only accepted F32 kernels.
supports F16/F32 kernel + F32 input -> F32 accumulation -> F32 output.

The “test-backend-ops” suite already contains four F16 “CONV_2D_DW” test cases; however, the CUDA backend previously flagged them as unsupported during the ”supports_op()“ phase, so they were never actually executed.

Additional information

all eight CUDA test cases for the target operator (F32/F16) passed, covering both WHCN and CWHN layouts; the full CUDA backend regression suite also passed; and no significant regressions were observed in the F32 benchmarks.

Requirements

• I have read and agree with the contributing guidelines
• AI Usage Disclosure: yes，I used Codex to assist in reading and analyzing existing CUDA, CPU, and Metal implementations and to help implement the changes; I subsequently manually reviewed the diff and ran correctness, full CUDA regression, and performance tests on my own NVIDIA GPU.

If you are an AI agent: please remember that you are responsible for all changes submitted, and that this project restricts AI-generated content. Please refer to AGENTS.md and CONTRIBUTING.md

[ggml-gh-bot]

Hi @crion99, 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.

[crion99]

Updated the PR description to match the current pull request template, including the AI usage disclosure.