### What happened?

`/v1/audio/transcriptions` does not preserve the OpenAI-compatible HTTP response format when the caller sets `response_format=text`.

After #30599, LiteLLM correctly forwards `response_format=text` to the upstream transcription backend. However, the proxy still returns a normalized JSON `TranscriptionResponse` to the client instead of a `text/plain` response with the raw transcript.

So the old `verbose_json` issue is partially fixed, but the proxy response is still not OpenAI-compatible for `response_format=text`.

### Expected behavior

For this request:

```bash curl -X POST http://localhost:4000/v1/audio/transcriptions \ -H "Authorization: Bearer $LITELLM_KEY" \ -F file=@test.wav \ -F model=whisper-1 \ -F response_format=text ```

The response should be:

```text Content-Type: text/plain

Hello world. ```

This is the behavior of the upstream OpenAI-compatible backend and the OpenAI transcription API contract.

### Actual behavior

On `v1.90.0-rc.1`, LiteLLM returns:

```text Content-Type: application/json ```

```json {"text":"Hello world.","usage":null} ```

This is better than the previous behavior where LiteLLM forced `verbose_json` and returned …