For advanced implementations, the r0 can also be used without considering register dependencies since it discards all writes and always read 0, so your "xor r0,r0" would possibly stall a long pipeline if the preceding instruction depended on the old value of r0 in a calculation before clearing it with xor.
Lastly, I guess one has to see it in the light of its day, loading registers with 0 was and probably is rather common at loop starts and so on, so it was deemed useful, just like some FPUs have instructions to read constants like PI, e and ln(2) just because it seems to be a good thing to waste a certain amount of transistors on if it is commonly used.
but the same advanced implementation that can break dependencies when reading/writing to r0 can do the same when handling xor r0,r0. Indeed that's exactly what many CPUs do.
I'm not qualified to say whether a zero register is better or not, but I will mention that it basically requires your ISA to be 3 operands that has a cost in term of instruction length: so you save bits to have a smaller set of instructions, but need more bits to increase the number of registers.
RISCs are usually 3 operands anyway for many reasons, so in practice it doesn't make a difference I guess.
xor r0,r0 seems a small price to pay.