fix: preserve reference image dimensions in server requests

[mikemikimike]

Summary

• Preserve the source dimensions of ref_images, sdapi extra_images, and OpenAI edit uploads during decoding.
• Keep output-size decoding for init and mask images unchanged.

Issue: #2004
Fixes #2004

Testing

• cmake -S . -B build -DSD_BUILD_EXAMPLES=ON -DSD_SERVER_BUILD_FRONTEND=OFF -DSD_WEBP=OFF -DSD_WEBM=OFF -DCMAKE_BUILD_TYPE=Release
• cmake --build build --parallel 4
• sd-cli --help
• sd-server --help

[leejet]

Please do not commit any test-related files.

[mikemikimike]

@leejet The current head is 1a6d542b6df1d4f3b70eedb27ee574e3b52ecff9. gh pr diff --name-only contains only examples/common/common.cpp, examples/server/routes_openai.cpp, and examples/server/routes_sdapi.cpp; the previously added test files are no longer present. No code changes were made in this pass. Please re-review.

[leejet]

The first reference image is also copied into init_image below. With this change, that copy retains the original dimensions and is later stretched to the output size by sd_image_to_tensor(), while the mask still goes through center-cropping and resizing during decoding.

When the uploaded image and mask have a different aspect ratio from the requested output, they therefore become misaligned, causing inpainting to affect the wrong region. I reproduced this with a 128×64 image/mask pair and a 64×64 output: they align before this change but not afterward.

Please keep the reference at its original size while separately applying the existing output-size crop/resize preprocessing to init_image, so it stays aligned with the mask.