Digital gain is free, sometimes a lower gain setting in a recording path is quieter than a higher one. I'd be more worried about maxing signal levels if I were transferring out of a recorder in the analog domain.
24 versus 16 is a real world difference of 25-30dB dynamic range, depending on the converter. Recording at -20dBFS in 24 is still more dynamic range than 0dBFS in 16.....and ya can't successfully record at 0dBFS, hardly at -6dBFS in unknown situations. Bring -20 up to -1 with digital gain in post, still more dynamic range than the 16 capture.
16 versus 24 within a 32float scenario; it's about what the 32float does to the bottom bit of the 16, which is 8 higher than the bottom bit of the 24. It's where the dither and quantization error lives. That's further buried in the 24, and therefore is also after processing. If you get into forensic fixes and noise reduction with something like RX7, there's a lot more in a 24 bit capture for the algorithms to work with. It keeps the artifacts of the medium much further out of the way, regardless of the dynamic range of the intended capture.