feat(gemma4): serve image input on the Gemma-4 releases

[jason-fxz]

Image input for the Gemma-4 releases: the gemma4 26B-A4B MoE and 31B dense through their ViT tower, and the 12B gemma4_unified release through its linear patch embedder, on the multimodal path from #454.

What Gemma-4 needs beyond the Qwen path

• Processor. <|image|> expands to <start_of_image>, the soft-token pad ids and <end_of_image>; 1-D rope. The count is one of the checkpoint's fixed budgets (70 / 140 / 280 / 560 / 1120, default 280): --image-max-tokens picks the largest budget within it and refuses one below 70 at start-up; --image-min-tokens has no effect.
• Tower. SigLIP-style ViT, 3x3 average pooling and the multimodal embedder on the text stack's projection layers, block weights host-streamed by default; the reader skips the tower tensors when no encoder is built.
• Attention. Per the checkpoint's use_bidirectional_attention: "vision", the sliding-window layers attend in both directions inside an image span; the full-attention layers stay causal. The scheduler hands the batch the end of each token's image span (Batch.mm_block_ends), the Triton extend kernels honor it within the window, the other backends never receive it.
• Chunking. On such a model a prefill chunk that would end inside an image span ends at the span start, decided after every other cap; a request whose image does not fit the budget left by other requests in the pass waits for a pass of its own. Only an image larger than a whole pass is split, with a one-time warning. Causal families keep the plain chunking.
• 12B gemma4_unified. No tower: a linear patch embedder (LayerNorm, dense 6912 -> 3840, LayerNorm, factorized 2-D position embedding, LayerNorm) feeds the same multimodal embedder, one 48x48 super-patch per soft token, about 95 MiB resident. Same processor call and budgets, same bidirectional spans. Adds freetoken.layers.LayerNorm next to the RMSNorm family.

Flags

No new flags. docs/cli.md describes the multimodal flags in the engine's own terms; docs/models.md carries the per-family image-input table.

Verification

• Tower and 12B embedder against the HF reference on real weights: mean cosine above 0.99 (tower, bf16 error within 1.2x of the reference's own noise) and 0.99999 (embedder).
• Triton extend kernels with block_ends match a reference attention under the same mask, plain and split-KV; scheduler tests cover the span pull-back, page alignment, the shared-pass deferral and the sliding-window cap.
• Greedy image smokes describe the synthetic image correctly on Gemma-4-26B-A4B-NVFP4 (MoE offload, both encoder placements, --image-max-tokens 100), Gemma-4-31B-IT-NVFP4 and gemma-4-12B-it; --text-model-only unchanged. At --max-extend-tokens 200 the 26B logs the cut of its image once and still describes it.
• Tokenizer tests on the real 26B and 12B checkpoints; full non-slow suite on the squashed tree 1800 passed (plus the four known environment failures on this box).

Not in this PR

• Video and audio input, including the 12B release's embed_audio.