Fix size parameter types and improve resize_image interpolation

[LexSong]

This PR fixes two issues in the resize_image function:

1. Make sure size parameters to be integers

   • The function sometimes got width and height as non-integer types, (like Tensor), which caused crashes when calling cv2.resize()
   • I added int() to make sure they’re always python integers. This stops the errors.

2. Improve interpolation logic

   • The old code used 'lanczos' when shrinking images and 'area' when not, which was backwards.
   • Now, it uses 'area' only when the image shrinks a lot (to half size or smaller), since 'area' is better for big reductions. Otherwise, it uses 'lanczos'.

[kohya-ss]

Thank you for this PR! The 1. makes sense.

Regarding 2., after a quick comparison, I found that PIL's LANCZOS seems to produce almost the same results as CV2's AREA when downsampling.

Original image (512x512):

Resized with AREA (320x320):

Resized with LANCZOS4, PIL (320x320):

However, resizing with PIL is about 10 times slower because it involves type conversion.

Therefore, if the resized image size is smaller than the original size, it is sufficient to simply use CV2's INTER_AREA.

[LexSong]

I tried with 90% downscale and there is no obvious quality difference. As you said, let's simply use AREA for all downscaling.

May I ask why we're using PIL's lanczos not CV2's lanczos?

[kohya-ss]

The update is done in #1426.

When upsapling128x128 to 328x328 (328 is an arbitrary choice),

CV2 LANCZOS:

PIL LANCZOS:

For example, you can see four gray dots where the lines intersect in CV2 image.

[LexSong]

I see, thanks for the reply. It's possible caused by gamma correction step. which I think it's skipped in opencv implementation.