coverage: test the Qwen3.5 GDN layer loader (extract_linear_layer_generic) #5214

Coverage gap from the nightly coverage report (#4458): SafetensorsToAprConverter::extract_linear_layer_generic (crates/aprender-serve/src/safetensors_infer_convert.rs) has 138 uncovered lines and 0% coverage. It loads one Gated Delta Net (linear attention) layer for Qwen3.5 SafeTensors models.

Nothing tests its two weight-naming conventions:

• combined self_attn.in_proj_qkvz + in_proj_ba, sliced at key_dim/value_dim boundaries;
• separate linear_attn.in_proj_qkv + in_proj_z/a/b.

Nothing tests its size-mismatch refusals, or its no-dense-FFN (MoE) branch either. A wrong slice offset there would hand the recurrence the wrong weights, and no unit test would notice.

done_when: unit tests drive both conventions through an in-memory tensor source and assert the exact slices. Both size-mismatch paths return errors. The MoE branch yields no gate and zero-filled up/down. A mutant on a slice boundary fails a test, and the gated-delta-net contract gains a falsification row for the weight split.