vulkan: fuse qwen4exp's SCALE -> SIGMOID -> SCALE -> hc_post chain

[fxgsell]

The PR is a Vulkan-only fusion that runs qwen4exp's SCALE -> SIGMOID -> SCALE -> hc_post chain as one kernel and stops the reorder pass from splitting it; it gives exact output and +3.8% decode, passes all 18,993 tests.

Overview

qwen4exp (Qwen3.8-Flash-Next) builds each hyper-connection's scatter weights as w = 2*sigmoid(inject/hc): three tiny ops on a 4-element tensor, right before dsv4_hc_post. That happens twice per layer, 96 times per token. On Vulkan, decode is limited by the number of kernel dispatches, not by arithmetic, so those 288 dispatches per token cost real time.

The change, all inside the Vulkan backend:

Part	What it does
New fusion HC_POST_GATE	Detects SCALE -> SIGMOID -> SCALE -> DSV4_HC_POST and runs it as one dispatch; hc_post computes the gate while loading post[] into shared memory
Match check	Fuses only when the op is sigmoid, neither scale has a bias, the input is f32, and the shapes match; anything else runs unfused
Reorder registration	Without it, Vulkan's reorder pass split all 96 chains per token (0 left adjacent), so the fusion could never fire
Allocation guard	add_pattern_alloc_deps keeps the input alive through the fused output, so the overlap check doesn't silently cancel the fusion
Test	A gated variable on the existing test_dsv4_hc_post; no new test file

Checks:

Check	Result
KL divergence vs master (40 x 512 tokens)	0.000000 (max 5.8e-5), same top token 100%
Decode, Flash-Next UD-IQ1_M	49.84 → 51.72 t/s (+3.8%)
Decode at 16k context	42.87 → 44.27 t/s (+3.3%)
hc_post op tests, R9700 and 7900 XT	9/9 each
Full test-backend-ops on Vulkan0	18,993/18,993 (master 18,990/18,990)
Prefill, 32k and 64k context	no change

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, the code, tests and benchmarks were AI-written. Design/review by human.

[ggml-gh-bot]

Hi @fxgsell, thanks for your contribution!

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

• 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.