Caps between +V and -V rails will only decouple/smooth for devices that are being referenced to one of those rails rather than 0V (like an opamp between the positive and negative rail with no third pin dedicated to ground).
Caps from either +V or -V to 0V most effectively decouples devices are referenced to 0V. Although caps in this manner will decouple devices connected between +V and -V with no connection to GND, they will not be as effective at this as they otherwise could be. This is because you could arrange those same capacitors in parallel between the +V and -V rail to achieve an effective capacitance four times larger, but the price you pay is that those caps will no longer decouple for devices referenced to 0V.
NOTE: Some bipolar voltage devices, such as comparators or opamps, with digital inputs for auxilliary functions may have a ground pin that is only there to provide a reference potential for the logic digital inputs (so you don't have to do something weird like connect your MCU ground pin to -V so it can drive the logic input), and is not involved with the return currents for the bipolar voltage output. These do not count as a ground referenced devices as far as the return currents at the output are concerned.
For something like bipolar output switching converters, you may need a cap between +/-V to 0V even if your load is only connected across +V and -V because if the converter operates with two halves (as would likely be the case if it were regulated or not isolated), both halves would need independent smoothing. If the architecture doesn't operate as two halves (i.e. unregulated converter with a tapped transformer output) then it possibly won't have that same need.
It should be evident by now that it is just safer to connect between each power rail to 0V. You cover more scenarios that way, especially on a supply where the load is unknown.