Theoretically, the correct filter is the one that has no attenuation up to half the sampling frequency and infinite attenuation above, ie. a brick wall filter, while also maintaining a constant delay across all frequencies, ie. phase linear. Only this filter will correctly reconstruct a properly band-limited sampled signal, ie. a signal that has no content above half the sample rate. Obviously, this filter would need to be infinitely steep which is practically impossible.
However, if there is out of band content in the sampled signal, ie. the signal is not properly band-limited, the filter has to remove (attenuate in practical terms) that content which results in the ringing eg. as that seen in the single sample impulse responses. The more the filter attenuates the out of band content, the more ringing there will be.
However, even if the signal is not properly band-limited, the frequency of the ringing is half the sample rate which is at least 22.05 kHz and thus of itself is most likely not audible. It may however result in audible intermodulation distortion. Less steep filters on the other hand result in less ringing but regardless of whether there is out of band content contained in the signal result in more imaging/aliasing which of itself may be audible.
The same holds for filters that are not phase linear such as the 'SD' filters. Thus, if the signal is properly band-limited there will be no ringing regardless of the selected filter and if the signal is not properly band-limited but there is no ringing there will be aliasing/imaging.
However, the steeper the filter and the more linear its phase response, the greater overall delay (latency) it introduces. Thus, if a steep filter and low overall delay (latency) is desired, the linearity of the phase response has to be compromised. As RME is (for the most part) a pro audio company, this might be the reason 'SD Sharp' is selected as the default filter.
The most accurate filter is 'Sharp' having the most linear frequency response up to 20 kHz and the highest attenuation of content above the Nyquist frequency, ie. lowest imaging/aliasing, while also being phase linear, ie. the delay is constant for all frequencies. If lower latency is desired, the 'SD Sharp' filter has very similar frequency response and similarly low imaging/aliasing but results in phase distortion which means that the delay is frequency dependant.
The 'Slow' and 'SD Slow' filters have lower latency than their sharp counterparts but this comes at the cost of a high frequency roll-off and less attenuation of content above the Nyquist frequency, ie. more aliasing/imaging.
The 'NOS' filter has the lowest latency of all filters while also being phase linear but has the highest high frequency roll-off and no significant attenuation of content above the Nyquist frequency, ie. a lot aliasing/imaging.
Boosting the high end of the slow and 'NOS' filters gives a more extended high end but still has a lot of aliasing/imaging while ringing is not an issue with properly band-limited content.
As mentioned by ramses and Curt962, if you want to test which filter you prefer, make sure that you have proper controls in place. This requires that you do not know which filter is selected during testing. If you prefer the high frequency roll-off of the slow or NOS filters, this can be implemented with the PEQ while avoiding aliasing.
TLDR: If the content is properly band-limited and latency is not a concern, the best filter is 'Sharp'. If low latency is desired and the content is properly band-limited, the best filter is 'SD Sharp'. If the content is not properly band-limited, determine for yourself whatever compromise of ringing, high frequency extension and aliasing you prefer.