The typically accepted "fair comparison" among two models is to give them roughly the same number of parameters, so yes, doubling the size of the decoder (where size is measured in #parameters) sounds like a reasonable way to accomplish this.

Is the capacity of a multitask U-Net with two-decoders the same of a standard U-Net with doubled capacity in the decoder?

It's kind of hard to quantify model capacity, since it's not directly related to the number of parameters -- if you stack a huge number of linear layers, that doesn't increase model capacity at all compared to a single linear layer.