[inureyes]

Summary

Implements #2794: five GELU formula-rule gaps outside #2548's row list, following the section-0b rule (the checkpoint config decides; an omitted key falls back to the Hugging Face config-class default; mlxcel follows HF where mlx-lm or mlx-vlm departs).

What changed

• GPT-BigCode reads activation_function (absent = the GPTBigCodeConfig default gelu_pytorch_tanh), resolves it once at load onto MLP, and rejects non-GELU names in validate()
• Idefics2 and Granite 4 Vision pass use_fast_gelu = false; the tower config decides (Idefics2 resolves its omitted key to the Idefics2VisionConfig default via default_missing_hidden_act_to_siglip)
• Shared VisionHiddenActivation maps every ACT2FN GELU name through GeluActivation::from_hf_name (gelu_new / gelu_fast now tanh) and rejects relu / silu as load errors; Bunny's sparse-config fallback propagates instead of swallowing
• A declared gelu_pytorch_tanh on the shared SigLIP MLP runs the compiled kernel like the omitted-key path (VisionMlpActivation::PytorchTanh removed, SiglipDefaultTanh folded in)
• Dead C++ compiled_gelu removed with its three FFI tests and both compile-audit entries; the XLA llava vision config accepts the same GELU name set
• docs/gelu-call-sites.md: rows T5, S6, S7, a new S-cfg row, the shared-resolution paragraph, the inventory rows, and the compiled_gelu row updated; GeluActivation "Used by" doc comment extended

Test plan

• cargo fmt --all -- --check passes
• cargo clippy --profile test-fast --features cuda -p mlxcel -p mlxcel-core -p mlxcel-xla --all-targets -- -D warnings passes (exit 0)
• cargo check (test-fast, cuda, the three packages) clean in 11m37s
• cargo test -p mlxcel-xla vision_config: 10 passed, 0 failed
• GPU lib tests (gpt_bigcode, siglip, vision::config, idefics2, granite4, llava, core gelu) with --test-threads=1 and MLX_ENABLE_TF32=0 under gpu-lock (run after the push, when the lock freed): 111 passed, 0 failed, 1 ignored on CUDA (GB10)
• VLM smoke (greedy mlxcel generate --image + one mlxcel-server image request each on models/idefics2-8b-4bit, models/granite-4.0-3b-vision-4bit, models/smolvlm-instruct-bf16, MLX_ENABLE_TF32=0, run after the push, on tests/fixtures/test_image.png): all six coherent. generate: idefics2 "A bright orange colored fabric with no patterns or text."; granite4 "The image provided is a solid block of a uniform color, specifically a bright orange or reddish-orange hue."; smolvlm "A solid orange background with no other objects, text, or scenery.". server: idefics2 "The image is a bright orange background with no other objects in the image."; granite4 "The image is a solid, uniform block of a bright orange color."; smolvlm "A solid orange background with no other objects, text, or elements present."
• GPT-BigCode smoke: not run, no checkpoint exists on this host (as the issue notes); the Metal benches use models/mlx/gpt_bigcode-santacoder on a Mac host

Closes #2794