textures: (un)swizzle large images in bands on the transcode workers

Texture uploads and the software 2D blitter convert between block-linear
and linear layouts on the GPU thread, one row at a time on one core
(~1.4 GB/s). In Mario Kart World's first minute that held the GPU thread
for up to 340 ms at a time (an 8192x8192 BC7 texture of 341 MiB), and
for 15-80 ms per blit through dozens of RGBA32F 2D blits into its
terrain layers while a course loads.

Swizzle and the subrect copies now split images of 4 MiB or more into
bands of whole block rows (about 1 MiB each), which read and write
bytes of their own on both sides, and run them with ParallelFor: the
calling thread and whichever transcode workers are free take bands
until none are left, so a pool busy with asynchronous decodes never
makes a copy slower than before.

Mario Kart World, first 45 s: the 341 MiB upload 340 -> 187 ms, the
144 MiB one 106 -> 75 ms, 37 software blits 857 -> 623 ms in all;
frames render as before.

0 files changed, 0 insertions, 0 deletions