I understood everywhere it would be best to use power-of-two-sized images
Indeed, but the power-of-two thing (non-necessarily square, btw) applies to whole textures, i.e. as loaded in the video memory ; due to data alignment and hardware architectures, these textures are faster to process.
When putting smaller textures/sprites into larger atlases (or spritesheets), the source size doesn't matter much, as long as the batching tool is smart enough to minimise wasted space and still produce power-of-two output sheets.
Actually sprites in a sprite sheets don't have to be aligned to anything at all, as long as textures coordinates match. Some atlasing tools are able to generate non-uniform grids, i.e. mixing sprites of different sizes efficiently.
It would be nice with C2 to at least be able to disable the 1px gap when using point sampling.