CUDA : use mmvf for tiny-N F32/F16 weights at decode batch sizes

[douyamv]

Small F32/F16 weight matrices with N ≤ 64 rows (hyper-connection inject/gate projections [K, 4], GDN β/α
projections [2560, 32] in qwen4exp) were dispatched to cuBLAS for decode-sized batches: a TF32 GEMM plus a
split-K reduction kernel, ~8 µs and two launches per call, 4 calls per layer. ggml_cuda_should_use_mmvf already
accepted N ≤ 16 for these types; this raises the limit to 64. The vector kernel does them in one launch with full
FP32 accumulation and is not slower for any of the shapes we measured on sm_80.

🤖 Generated with Claude Code

[ggml-gh-bot]

Hi @douyamv, 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.

• Multiple open PRs from a new contributor: We limit new contributors (those without a previously merged PR) to 1 open PR at a time. You currently have 5 open PRs.

• AI-generated content: While code is allowed to be generated by AI, please write the PR description and commit messages on your own without the help of AI.

---

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

[JohannesGaessler]

According to the llama.cpp AI usage policy:

It is strictly prohibited to use AI to write your posts for you (bug reports, feature requests, pull request descriptions, Github discussions, responding to humans, ...).