There are several reasons actually. One is that GPUs have different memory tradeoffs than CPUs and prefer higher bandwidth at the cost of latency; unified memory is always a compromise, usually in favor of the CPU. Another is that top discrete GPUs are reticle limited, which means it's not possible to make the chip any bigger, so adding other SoC parts to the same chip would necessarily reduce the space dedicated to graphics. Yet another is that separating the GPU and CPU makes it easier to effectively cool both of them.