Yes, all four: the views show the object's front, both sides and back, and none of them is from the input image's own viewpoint. Idol generates the splat from that single image rather than training it on views, so the sides the image does not show are generated too.
You are right that the number measures agreement with the reference. At every view, the back included, we compare the phone's splat of 65,536 Gaussians with the splat of 262,144 that the author's PyTorch pipeline generates from the same input image and the same random numbers, both rendered by the same renderer from the same cameras. The back view agrees at 33.04–35.05 dB over the three test inputs that gave the 32–34 dB. Fidelity to the input image, or to the real object's back, was not measured.