fix(gemma3): apply rope_scaling to global attention layers

[1qh]

Problem

Gemma 3 ships rope_scaling in text_config and applies it to the global-attention layers only:

"rope_scaling": {"factor": 8.0, "rope_type": "linear"}

TextConfig declares the field, but Attention.__init__ builds a plain nn.RoPE and never passes it, so global layers run unscaled positions (effective scale=1.0 where the model expects 1/8).

The error is negligible at short context and compounds with length. On mlx-community/gemma-3-4b-it-4bit, first-token logits for a 2916-token prompt (mlx-vlm vs the reference mlx_lm.models.gemma3):

token	mlx-lm (reference)	mlx-vlm (before)
Từ (correct continuation)	33.75 — argmax	35.00 — 2nd
Đối	31.62 — 2nd	36.25 — argmax

Greedy decoding therefore picks a different token and degenerates into a repetition loop, while the same weights under mlx_lm.generate answer correctly. Short prompts are unaffected (top-100 logit rank-correlation 0.96 at 195 tokens vs 0.53 at 2916), which is why this hides in ordinary use.

Fix

Use the existing rope_utils.initialize_rope, matching gemma4 in this repo and gemma3_text in mlx-lm. Sliding layers keep unscaled local RoPE; scaling applies to global layers only.

After the fix, mlx-vlm reproduces the reference logits exactly:

top5: [(33.75, 'Từ'), (31.62, 'Đối'), (29.38, 'T'), (25.75, 'From'), (25.62, 'Với')]   # mlx-vlm, patched
top5: [(33.75, 'Từ'), (31.62, 'Đối'), (29.38, 'T'), (25.75, 'From'), (25.62, 'Với')]   # mlx-lm reference

Scope

• gemma3 — the reported defect; gemma3_text inherits the fix through gemma3.language.LanguageModel.
• gemma3n — same shape (declares rope_scaling, built a plain nn.RoPE). With no rope_scaling in the config the behaviour is unchanged (scale=1.0), so this is a no-op for current checkpoints and correct if one declares scaling.

Tests

Two regression tests added to test_models.py:

• test_gemma3_rope_scaling_applies_to_global_layers — asserts global layers get 1/factor, sliding layers stay 1.0. Fails on main with AssertionError: 1.0 != 0.125.
• test_gemma3_without_rope_scaling_is_unscaled — asserts the no-scaling path is untouched.

python -m unittest mlx_vlm.tests.test_models passes (all 20 gemma tests included).

[1qh]

Follow-up: I swept the rest of the repo for the same shape and found ten more models that declare rope_scaling and build a plain nn.RoPE without it — including mllama, which affects Llama-3.2-Vision today. That is #1710.

The two are the same defect class and can be reviewed together or merged in either order; they touch disjoint files. This one is separate because it is verified end to end against a real gemma-3 checkpoint (patched mlx-vlm reproduces mlx-lm's reference logits exactly and a served repetition loop disappears), whereas #1710's evidence for mllama is a numerical comparison of the required Llama3RoPE against the current plain RoPE rather than a full model run. Happy to squash them into one PR if you would rather review a single change.

[Blaizzy]

LGTM, thanks!

[lucasnewman]

@1qh Can you sign your commits and force push so we can clear the merge requirements?