For anything GPU-rendered, applying a rotation matrix to a texture sample and/or frame-buffer write is trivially cheap (see also why Vulkan prerotation exists on Android). Even ignoring GPU-rendering, you always are doing a copy as part of rendering and often have some sort of matrix operation anyway at which point concatenating a rotation matrix often doesn't change much of anything.
The cost is paid in different memory access patterns which may or may not be mitigated by the GPU scheduler. It's an insignificant cost either way though, both for the encoder and the renderer. Also depending on the pixel order in sensor, file or frame buffer "rotated" might actually be the native way and the default is where things get flipped around from source to destination.
Access pattern is mitigated by texture swizzling which will happen regardless of how it's ultimately rendered. So even if drawn with an identity matrix you're still "paying" for it regardless just due to the underlying texture layout. GPUs can sample from linear textures, but often it comes with a significant performance penalty unless you stay on a specific, and undefined, path.
