Fix Gemini media prompt sanitization

[zy6p]

Summary

• sanitize Gemini image/video prompts before forwarding them to Flow
• ignore tool/function-calling scaffolding in systemInstruction for media models
• strip <tools>...</tools> blocks and agent preamble lines from media prompts

Why

Gemini-compatible clients sometimes send agent/tool scaffolding such as:

• You are a function calling AI model
• <tools>...</tools> blocks
• JSON schema for tool definitions

For text models that can be acceptable context, but for Flow media generation these lines are not valid prompt content. flow2api currently forwards them verbatim, which can make image generation fail with:

PUBLIC_ERROR_UNSAFE_GENERATION: Request contains an invalid argument.

This is reproducible even when the actual image prompt is valid. The failure is in the route normalization layer, not the Flow model mapping itself.

Verification

• reproduced success for clean gemini-3.1-flash-image-square and gemini-3.1-flash-image-four-three prompts
• reproduced failure when the same request included agent/tool scaffolding in the media prompt
• after this patch, the normalized media prompt only keeps the actual image description and generation succeeds