Personally I always stick to multiple of 8 but here it is:
Multiple of 4: In H.265, DCT can be done in 4x4 Transform Units, so it's not a problem for that. It can also use motion vectors for 8x4 or 4x8 Prediction Units if there is only a single reference frame, or 8x8 if there are 2 reference frames. I suspect this causes some inefficiency, but nowhere near as bad as not-divisible-by-16 was for MPEG-4.
Multiple of 2: I would not do that. Both PUs and TUs have to be padded. (I suspect it's still better than MPEG-4, just because the "bad" blocks are not that huge.) Also, what are you gaining? Cutting off 2 pixels is not going to be noticeable.
Multiple of 1: You can't even do it unless the pixel format is YUV444. Don't do it unless it's a really low resolution video, like maybe you are using it as a replacement for GIFs.