Fix batched causal mask (b>1) across Qwen3/GLM4/SmolLM3/Z-Image serving models

[astorise]

Fixes the batched causal-mask under-sizing bug across the fork's serving-optimized models, by cherry-picking upstream huggingface#3780 (@aashish-thapa, addresses huggingface#3582). Commits and authorship preserved.

The bug

8 models build the causal mask as Tensor::from_slice(&mask, (b, 1, tgt, tgt + offset), …) while the slice only holds tgt * (tgt + offset) elements — one batch element's worth. For b > 1 this under-sizes the storage: from_slice against a concrete shape skips the length check, so the GPU standard-attention path (l > 1 prefill) gets garbage/NaN, and CPU-flash builds fail outright. Affected: qwen3, quantized_qwen3, qwen3_moe, quantized_qwen3_moe, glm4_new, quantized_glm4, smol/smollm3, z_image/text_encoder — i.e. exactly the batched-serving path.

The fix

Build the mask batch-independently as (1, 1, tgt, tgt + offset) and let broadcast_add fan it across the batch — correct for any b, and cheaper than materializing b copies. Includes a CPU regression test (qwen3::tests::causal_mask_is_batch_independent) asserting the shape, the causal-triangle values, and that broadcasting onto b > 1 scores gives every row the same mask.

Why huggingface#3780 (arbitration among the cluster)

Several upstream PRs target this "Qwen3 mask for batch>1" family (huggingface#3646/huggingface#3624/huggingface#3610/huggingface#3586/huggingface#3611/huggingface#3609/huggingface#3484/huggingface#3743). huggingface#3780 is the right pick for this fork:

• It targets the fork's exact code — the same causal_mask helper across the same 8 files — and cherry-picks cleanly, because the fork inherited that helper from upstream.
• It's the minimal broadcast-correct fix, consistent with the fork's serving-optimized style, and ships a CPU regression test.
• Fix attention masks for chunked prefill (seq_len > 1 with a warm KV cache) huggingface/candle#3743 fixes a different scenario (chunked prefill with a warm KV cache; touches falcon/olmo/phi/qwen2 — none of the 8 affected models). Fix causal mask construction for batch size > 1 in Qwen3/GLM4/SmolLM3/Z-Image huggingface/candle#3484 conflicts and is messier. The rest are single-model subsets.

CI gate (candle's own, all green)

• cargo fmt --all -- --check
• cargo check --workspace
• cargo clippy --workspace --tests --examples --benches -- -D warnings
• cargo test -p candle-transformers --lib causal_mask — regression + sibling mask tests pass