fix(inference): support string URL output in Replicate text-to-image (#1835)

This PR adds support for Replicate text-to-image models that return a
single URL string instead of an array of URLs.

This should fix
replicate/huggingface-model-mappings#31, which
is currently failing:

<img width="1048" height="432" alt="image"
src="https://github.com/user-attachments/assets/a9750364-7b8d-4f2f-96e9-4a1afb39af40"
/>

## Solution

The `ReplicateTextToImageTask.getResponse` method previously only
handled array outputs, causing errors for models that return a single
string URL. This change makes the text-to-image implementation
consistent with other tasks like `ReplicateImageToImageTask` and
`ReplicateTextToVideoTask`, which already support both formats.

## Prompts

> Does the replicate text-to-image implementation in this codebase
support models that return a URL as output? Or does it always expect an
array of image URLs?

> Let's update it to allow URL strings as a valid output response

> commit the changes to a branch, push it, and open a PR against the
upstream repo: huggingface/huggingface.js

---------

Co-authored-by: Celina Hanouti <hanouticelina@gmail.com>

Author: zeke

packages/inference/src/providers/replicate.ts

@@ -95,6 +95,20 @@ export class ReplicateTextToImageTask extends ReplicateTask implements TextToIma
 	): Promise<string | Blob | Record<string, unknown>> {
 		void url;
 		void headers;
+
+		// Handle string output
+		if (typeof res === "object" && "output" in res && typeof res.output === "string" && isUrl(res.output)) {
+			if (outputType === "json") {
+				return { ...res };
+			}
+			if (outputType === "url") {
+				return res.output;
+			}
+			const urlResponse = await fetch(res.output);
+			return await urlResponse.blob();
+		}
+
+		// Handle array output
 		if (
 			typeof res === "object" &&
 			"output" in res &&