fix(qwen3_vl): build causal mask for prefill, not single-token decode (#3505)

[0xSoftBoi]

Summary

In Qwen3VLModel::forward, the gate that decides whether to build a causal attention mask is inverted:

let attention_mask = if seqlen <= 1 {
    Some(self.prepare_decoder_attention_mask(...))   // mask built when *not* needed
} else {
    None                                             // mask skipped when needed
};

The decoder causal mask is only useful when more than one token is being processed at once (prefill / multi-token prompt). For single-token autoregressive decode steps the KV cache already constrains attention and the mask can be skipped. The current code does the opposite.

Every other candle text model (mistral, starcoder2, helium, olmo2, phi3, deepseek2, csm, gemma4, …) follows the canonical pattern, e.g. mistral.rs:

let attention_mask = if seq_len <= 1 {
    None
} else {
    let mask = self.prepare_decoder_attention_mask(seq_len, seqlen_offset)?;
    Some(mask)
};

This PR aligns Qwen3 VL with that shape and adds a comment explaining the invariant so the inversion isn't reintroduced.

Test

• cargo build -p candle-transformers — clean
• cargo clippy -p candle-transformers --lib -- -D warnings — clean
• cargo fmt --all -- --check — clean

Fixes #3505