fix: honor flash attention flag in LLM text encoder attention

[linxuhao]

Summary

LLMAttention::forward passed a hard-coded false for flash_attn to ggml_ext_attention_ext, so --fa never applied to LLM text encoders (Qwen, Mistral, Gemma, ...), unlike the diffusion models and VAEs that pass ctx->flash_attn_enabled.

Without it the [N*num_heads, n, n] attention matrix is materialized. On Vulkan devices limited to a 4 GiB storage buffer this crosses the limit at ~4096 tokens for 64-head encoders and the ops fall back to CPU (encode 11.5 s at 4065 tokens → 239 s at 4214 tokens on an RX 7900 XTX); on other backends it costs O(n²) memory.

This passes ctx->flash_attn_enabled, matching ernie_image.hpp. Behaviour without --fa is unchanged. The GPT-OSS path (attention sinks) is untouched.

Related Issue / Discussion

Fixes #1986

Additional Information

• Built sd-cli with -DSD_VULKAN=ON on master cc515a0 with this change.
• Measured with the same conditioner code on RX 7900 XTX (RADV, Mesa 25.2.8), Qwen3-VL-32B Q4_K_M: 4214 tokens 239 s → 9.7 s, 5081 tokens 357 s → 12.0 s, 15644 tokens 49.7 s; ≤4096-token outputs are identical when --fa is not set.

Checklist

• I have read and confirmed this PR follows the contribution guidelines.

AI-assisted: the investigation, measurements and change were done with Claude Code and reviewed by me.

🤖 Generated with Claude Code