kleidiai: fix work size and threads sync for fp16

[chaxu01]

This patch also updates the kleidiai lib to v1.14.0.

[chaxu01]

This PR addresses FP16 work-size and thread synchronization in the KleidiAI backend and also updates the KleidiAI library to v1.14.0. The changes only affect the KleidiAI path, which is gated behind GGML_CPU_KLEIDIAI and not enabled in the default CI builds. The two CI failures shown here are unrelated, as neither job includes KleidiAI in the build. The arm64 SVE job failed due to a rerank score outside the expected range, and the x64 AMX job failed with a ppl value above threshold along with a missing log file. I believe these are independent of this patch.

[ggerganov]

It would be useful to have a CI workflow that exercises the KleidiAI builds. Please consider adding one.

A starting point could be the new SVE workflow:

llama.cpp/.github/workflows/build.yml

Lines 1384 to 1408 in 5f7e166

	ggml-ci-arm64-cpu-high-perf-sve:
	runs-on: ubuntu-22.04-arm
	steps:
	- name: Clone
	id: checkout
	uses: actions/checkout@v4
	- name: ccache
	uses: ggml-org/[email protected]
	with:
	key: ggml-ci-arm64-cpu-high-perf-sve
	evict-old-files: 1d
	- name: Dependencies
	id: depends
	run: |
	sudo apt-get update
	sudo apt-get install build-essential libcurl4-openssl-dev
	- name: Test
	id: ggml-ci
	run: |
	LLAMA_ARG_THREADS=$(nproc) GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt

[chaxu01]

Thanks, @ggerganov — good suggestion. We’re definitely interested in adding CI coverage for KleidiAI builds. We’ll explore how best to integrate KleidiAI into the CI without adding too much overhead, and follow up with a separate PR to cover this.