Support resizing images smaller than 1024 during inference

Hi, thank you for your excellent work!

As I understand it, StrSR is trained and evaluated at a resolution of (1024 \times 1024). However, many commonly used SR benchmarks contain images whose dimensions are smaller than 1024. Directly running inference at resolutions that differ substantially from the training resolution may affect restoration performance.

Would it be possible to add an optional preprocessing strategy to the inference code, similar to SeeSR and OSEDiff? Specifically, if the bicubic-upscaled input is smaller than (1024 \times 1024), it could first be resized to (1024 \times 1024) for model inference, and then the restored result could be resized back to the original target resolution before saving.

Thank you!