Fix SGD image desc indexing for 3D UASTC HDR 6x6i textures

[SashaRX]

Fixes transcoding of 3D textures encoded to UASTC HDR 6x6 intermediate.

Problem

transcodeUastcHDR6x6_intermediate indexes the SGD image description table with

sgdImageDescIndex = level * levelImageCount + image

using the current level's image count. The encoder (basis_encode.cpp) writes the ktxUASTCHDR6x6IntermediateImageDesc entries in level order, level 0 first, with each level contributing numLayers × numFaces × depth(level) descriptions. The two agree only while every level has the same image count. For 3D textures depth halves with each mip level, so the transcoder selects descriptions belonging to other levels; since slice offsets are relative to their own level's data, transcoding fails with KTX_TRANSCODE_FAILED (or could silently read wrong slices when the offsets happen to stay in range).

This reproduces with libktx's own encoder output: encode a 24×24×8 VK_FORMAT_R16G16B16A16_SFLOAT texture with 4 mip levels via ktxTexture2_CompressBasisEx with KTX_BASIS_CODEC_UASTC_HDR_6x6_INTERMEDIATE, then ktxTexture2_TranscodeBasis(..., KTX_TTF_ASTC_HDR_6x6_RGBA, 0) fails. A 2D texture with otherwise identical parameters transcodes fine — presumably why this hasn't been noticed.

Fix

Accumulate per-level first-image indices before the level loop and index with firstImages[level] + image — the same approach transcodeEtc1s already uses with its firstImages table. No behavior change for textures where every level has the same image count (2D, arrays, cubemaps): for those firstImages[level] == level * levelImageCount.

Since firstImages[numLevels] is the image count implied by the texture's dimensions, the descriptor table size is now also validated up front: a table that is not exactly that many descriptions, or not a whole number of them, is rejected with KTX_FILE_DATA_ERROR before any level is processed, instead of surfacing as KTX_TRANSCODE_FAILED per image.

Tests

Adds generated 2D control and 3D multi-mip round-trip regression tests to transcodetests (in-process encode → transcode), with no new binary test resources. The 3D case fails without the fix and passes with it; both verify the post-transcode state (vkFormat, supercompression scheme, image data present).

Found while studying the transcoder for the per-level streaming work discussed in #1224.

[CLAassistant]

All committers have signed the CLA.

[MarkCallow]

Excellent catch. Thanks. As with PR #1236 , please sign the CLA. It should be sufficient to do it in just one PR.

[MarkCallow]

I won't run the workflows until you fix the incorrect for loop.

[SashaRX]

Done — switched the accumulation to the same 1..numLevels loop form transcodeEtc1s uses and applied your comment wording. Both regression tests still pass.