Why a bank of RAM for sound?
Because the SPC-700 is not just some DMA driven DAC unit, but a separate CPU working on its own 64 KiB address space. It's a greatly enhanced 6502 executing a program loaded from the main CPU to produce sound. In the most simple way this can be just a sample, but already here it can start/stop/chain and modify it according to messages send over by the main CPU. Or, on the other end of capabilities synthesize sound complete on its own.
I would've expected that if the sound chip is playing samples, it would fetch them directly from ROM where they are stored.
It's the same basic reason, why the Video system also has its own 64 KiB RAM. Performance.
Sure, copying each sound whenever played (and played and played) from ROM would work if it was only a simple DAC based system. But even then, fetching will collide with program (and data) fetches of the main CPU, thus dragging down system performance. Keep in mind, the SNES is still a classic 8 bit system with a rather meagre memory performance (1.8..3.5 MiB/s maximum bandwidth) (*1). Using a separate sound RAM is much like the use of sprites for such systems: saving on bandwidth of the main system. Only in this case we have a rather late design with comparable huge resources assigned to the task.
Was it customary to first copy them into the sound RAM? If so, why?
As said, it relieves the main CPU bus from data/program fetches of the sound CPU. Only one access during setup, and zero load thereafter. It can't get better, can it?
In fact, some games even double buffer the transfer by bringing sound data first into main RAM (as ROM may be slow) and then transfer it at high speed into sound RAM.
*1 - Nonetheless, with its PPUs it can be seen as a quite advanced one.