[Feature] Cache or reuse prompt embeddings across generate_image calls

[mayocream]

FLUX.2 Klein recomputes Qwen3 conditioning on every generate_image call, even when the prompt is unchanged. The current C API only accepts prompt strings, so bindings cannot cache and reuse c_crossattn/prompt embeddings.

Would you consider either:

• a bounded per-context cache keyed by prompt and active LoRAs, or
• an opaque encode/free condition API that generate_image can consume?

The cache should cover only text conditioning before request-specific c_concat/reference latents are attached. This would help repeated inpainting/edit calls using the same prompt with different images. Diffusers supports this through precomputed prompt_embeds.

[fffonion]

This will be useful for a preset prompt like I simply wants to "Fill the masked part of the image" etc, so that on user side they don't need to download and load the LLM every time.