Read attached chat images with a vision model (#82)

## Summary
- Image attachments (PNG, JPEG, WebP, GIF) skip the document parser and go to a vision-capable Ollama model (`OLLAMA_MODEL_VISION`, default `gemma4:latest`).
- The image is stored on the prompt. `PromptSerializer` returns `file_type` and base64 `image` so history can redraw it.
- Unsupported image types and files over 8 MB return a websocket error and are not saved.
- A later text turn does not reuse a previous image as a spreadsheet.

Pairs with `chat_web_app` `feat/128-vision-image-prompt` ([#128](ai_ml_operations/chat_web_app#128)).

## Test plan
- [ ] `manage.py test chat_backend.tests.test_vision chat_backend.tests.test_consumers.GraphNodeTestCase`
- [ ] Send a prompt plus PNG over `ws/chat_again/` and confirm the vision model is called.
- [ ] `GET conversation_details` includes `image` and `file_type` for that turn.
- [ ] CSV upload still follows the document path.

Reviewed-on: #82

.env.prod.example

@@ -37,6 +37,8 @@ OLLAMA_BASE_URL=http://10.0.0.128:11434
 OLLAMA_MODEL=gpt-oss:20b
 OLLAMA_MODEL_THINKING=gpt-oss:20b
 OLLAMA_MODEL_FAST=gemma4:latest
+# Vision (image + prompt). Must be a vision-capable tag.
+OLLAMA_MODEL_VISION=gemma4:latest
 OLLAMA_MODEL_UTILITY=llama3.2
 OLLAMA_EMBED_MODEL=nomic-embed-text
 OLLAMA_NUM_CTX_THINKING=16384