fix: honour vision=false by stripping image_url blocks before API call

[Veyal]

Problem

When a user sets vision: false on their chat model, Agent Zero still forwards image_url content blocks to the LLM provider. This causes providers that do not support image input (most text-only models on OpenRouter and others) to return:

{"error":{"message":"No endpoints found that support image input","code":404}}

The vision field on ModelConfig was defined but never consulted when building the message payload in LiteLLMChatWrapper._convert_messages. All message content, including images, was passed through images.prepare_content() unconditionally.

Fix

helpers/images.py — adds strip_images(content) which removes image_url blocks from a content list:

• Image-only content (no text) → returns "" instead of [] to avoid empty-content API errors
• Single remaining text block → collapsed from [{"type":"text","text":"..."}] back to a plain string
• Multiple remaining text blocks → returned as-is

models.py — _convert_messages now reads self.a0_model_conf.vision and calls strip_images() when vision is disabled.

Tests

7 new unit tests added to tests/test_vision_load_image_refs.py covering:

• String passthrough (no-op)
• Image-only list → ""
• Multi-image-only list → ""
• Mixed text+image → collapsed to plain string
• Multiple text blocks with image → list with images removed
• Text-only list → collapsed to plain string

All 7 tests pass.

Reproduction

1. Set chat model to any text-only model (e.g. deepseek/deepseek-r2) on OpenRouter
2. Disable vision in settings
3. Send a message — if any prior message in context contains an image attachment, the error No endpoints found that support image input is returned