GPU renderer double-copies locked textures

[slouken]

The GPU 2D renderer keeps an internal pixels buffer that it returns to the application in SDL_LockTexture(), and then allocates a GPU transfer buffer and copies the internal pixels to that, doing two copies for every texture upload.

Instead, it would be nice to allocate the transfer buffer in SDL_LockTexture() and let the application write directly to that and then upload to the GPU on unlock. If we do that we eliminate the double copy, don't need to keep a copy of the pixels around, and SDL_UpdateTexture() can be implemented in terms of LockTexture.