write ktx imageSize field as uint32 not size_t

[sahvx655-wq]

Noticed the .ktx files written by a -dl decode came out four bytes longer than the header plus surface should account for, which is what led me here.

1. the per-level imageSize is written with file.write(&image_bytes, sizeof(image_bytes)), and image_bytes is now a size_t since the write-buffer sizing was widened, so eight bytes are emitted.
2. the KTX1 container defines that field as a fixed 32-bit unsigned value, and load_ktx_uncompressed_image reads it straight back as a uint32_t, so the writer disagrees with both the format and our own loader.

The four stray bytes land in front of the surface, so a re-load picks the pixels up from the wrong offset and quietly corrupts the round trip rather than failing cleanly. Narrowed the value to a uint32_t for the serialised field and left the size_t sizing maths alone.