fix(shroud): Stop the shroud upload from exhausting memory on device reset

[tintinhamans]

The shroud was uploaded to the GPU with CopyRects on every frame. On NVIDIA the driver keeps a little memory for each of those calls and only allocates it when the device is reset. After a long match a single reset (lock screen, UAC prompt, alt-tab in fullscreen) allocates hundreds of MB in one go, and the 32-bit process runs out of address space. Reset then fails with out of memory, which is where crashes like CLIENT-55X come from.

The shroud now lives in a plain buffer, m_shroudData, and is copied into a managed texture through a lock. CopyRects, the system-memory surface it copied from and the pitch member are gone. This is the same change as upstream TheSuperHackers#3440 (issue TheSuperHackers#3439), and the same design BFME ended up with.

Measured on the upstream build, same skirmish at 120 fps for 10 minutes and then a forced device reset: main lost 316 MB of address space in that one reset, with this change it lost nothing. The upload is also cheaper, W3DShroud::render went from 0.032 ms to 0.006 ms per frame. On our own test setup an earlier version of this fix went through a 20 minute match and a screen lock without losing memory.

The second commit fixes the m_finalFogData fill in fillShroudData, which advanced the wrong pointer and only filled the first row. That code is behind DO_FOG_INTERPOLATION, so nothing changes in current builds.

Both GeneralsMD and Generals are changed. Zero Hour builds here; the Generals half could not be compiled in this tree, it is the same code that builds upstream.

[elopezgomez567-stack]

git init
git add .
git commit -m "Initial commit"
git push origin main

[elopezgomez567-stack]