perf(gemma4): fuse routed expert GELU gating

[pculaf]

Replace two tensor copies, GELU, and multiplication in Gemma 4's routed experts
with one ggml_geglu operation. This computes the same GELU(gate) * up result
using an existing kernel, reducing four GPU kernel executions to one per MoE
layer. For Gemma 4 26B-A4B, that removes 90 kernel executions per ordinary decode
step.

The change covers only routed experts. A shared-expert variant showed no
measurable benefit in preliminary testing, so it is excluded.

Tested on NVIDIA A10 with Gemma 4 26B-A4B IT Q4_K_M, a Q8_0 Gemma DFlash
draft, F16 KV and temperature 0. Prompt caching was disabled, and performance
timings were collected without a GPU profiler running.
Benchmark revision:
97d7c74.

The server's context capacity was 2048 tokens per request, covering input plus
generated output. Actual requests were shorter:

• Coding: all ten prompts from
  bench_he.py,
  sent through Gemma's chat template: 100–161 input tokens and 256 output tokens.
• Math: the first five cases from
  bench_gsm.jsonl,
  with 44–126 input tokens and a 1024-token output cap. Actual outputs were
  150–1024 tokens in ordinary mode and 195–1024 with DFlash.

Input counts include the chat template. Generation could end before its output cap;
the largest input-plus-output length observed was 1150 tokens.

Each mode used A/B/B/A order on one GPU (A: original, B: patched). Each session
included one full-corpus warmup pass followed by two measured passes, giving
four measurements per variant per prompt. Times below sum the per-prompt
median decode durations; speedup is original / patched - 1.

Mode	Prompts	Original (ms)	Patched (ms)	Speedup
Ordinary	Coding	29426.85	29362.55	+0.22%
Ordinary	Math	49301.40	48892.75	+0.84%
DFlash	Coding	15223.15	14928.10	+1.98%
DFlash	Math	16313.55	16114.85	+1.23%

DFlash gains were consistent across both A/B pairs. Ordinary results were
inconclusive: four coding prompts were slightly slower, and the math gain
varied from 1.40% to 0.09% between pairs.

All 240 timed responses and 120 warmups preserved the original outputs, token
counts and speculative acceptance within each mode.

[cubic-dev-ai]

No issues found across 1 file

_{Re-trigger cubic}