Fix RGB data handling in fetch_rgb_geotiff function

- Adjusted the order of method calls to ensure NaN values are filled after converting the dataset to float32.
- This change improves the robustness of RGB imagery processing in the inference pipeline.

infer.py

@@ -97,7 +97,7 @@ def fetch_rgb_geotiff(url: str) -> tuple[Image.Image, dict[str, Any]]:
                 raise ValueError(
                     f"GeoText requires RGB imagery; input has {dataset.count} band(s)."
                 )
-            rgb = dataset.read([1, 2, 3], masked=True).filled(np.nan).astype(np.float32)
+            rgb = dataset.read([1, 2, 3], masked=True).astype(np.float32).filled(np.nan)
             metadata = {
                 "crs": str(dataset.crs),
                 "bounds": list(dataset.bounds),