GeminiVideoOmni calls generateContent, but gemini-omni-flash-preview requires the Interactions API

## Summary

The built-in **Google Gemini Omni (Video)** node fails on Comfy Cloud before generation starts.

```
API Error: gemini-omni-flash-preview is only supported in the Interactions API and cannot be called directly via generateContent.
```

The workflow inputs are valid: the prompt is connected through the prompt socket, three reference images are connected, and the selected model is **Omni Flash**.

## Steps to reproduce

1. Open Comfy Cloud.
2. Add the built-in `Google Gemini Omni (Video)` / `GeminiVideoOmni` node.
3. Select `Omni Flash`.
4. Connect a non-empty prompt (reference images are optional; the reported case uses three).
5. Run the workflow.

## Actual result

The node immediately fails with the API error above. The node reports 0 seconds of server processing, so the request is rejected before video generation begins.

## Expected result

The node should call the Gemini **Interactions API** and return the generated video/audio plus any text response.

Google's Gemini Omni Flash documentation uses:

- SDK: `client.interactions.create(model="gemini-omni-flash-preview", ...)`
- REST: `POST https://generativelanguage.googleapis.com/v1beta/interactions`

Documentation: https://ai.google.dev/gemini-api/docs/omni

## Source-level root cause

In `comfy_api_nodes/nodes_gemini.py`, `GeminiVideoOmni.execute` currently submits:

```python
ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model_id}", method="POST")
```

with a `GeminiGenerateContentRequest` and `GeminiGenerateContentResponse`. That is the existing generateContent path used by other Gemini nodes, but `gemini-omni-flash-preview` explicitly rejects generateContent.

The node was introduced in #14695. Its PR checklist shows QA was not completed.

## Suggested fix

1. Add/use a Comfy proxy route for Gemini `v1beta/interactions`.
2. Serialize prompt/images/videos into the Interactions `input` schema.
3. Parse video output from the Interactions response `steps` array (or the SDK convenience `output_video` field where applicable).
4. Preserve current billing extraction and video download behavior.
5. Add a regression test asserting that `GeminiVideoOmni` does not use the generateContent endpoint/request model.

## Environment

- Product: Comfy Cloud
- Node: `GeminiVideoOmni`
- Model: `gemini-omni-flash-preview`
- Observed: 2026-07-17

Start in comfy_api_nodes/nodes_gemini.py at GeminiVideoOmni.execute and compare its current generateContent request with the Gemini Interactions API requirements. Trace the Comfy proxy route, input serialization, response steps, billing extraction, and video download behavior. Add a regression test confirming the node no longer uses the generateContent endpoint/request model and verify video, audio, and text outputs remain available.

Written by the indexing model from the issue text.