The bit depth of the files don't matter. Everything will be converted to float (32 or 64 bit, depending on your settings) in real-time before it enters the mix engine anyway. Hence the format of the mix-down will be float as well and needs dithering when exporting to integer (16 or 24 bit). Render setting (for intermediate bounces) should be at least 32 bit, in order to avoid unnecessary dither steps. (Before you ask: there's no dithering, going from 64 to 32)
 
From the above it should be clear, that dithering anything before mixing/processing doesn't make sense.