If the dithering is done right, dithering 64-bit double-precision to 32-bit single-precision floats doesn’t hurt. And if you’re archiving the audio to a file with 32-bit floats, it probably should be done, because the only cost is some processing. And MIPS are cheap in a non-realtime process.
It’s just that the quantization level is 150 dB below the loudest audio in the recording (so, even if you have 10 dB of headroom, with 32-bit floats, then the quantization level is -160 dBFS) and I sorta doubt that even bad quantization can be heard.
From a DSP programming perspective, it’s more of a pain-in-arse to dither and noise shape floating point, but it can be done. It’s just code and MIPS. May as well do it, if it’s not real time and no one is worrying about the MIPS.