There's been considerable discussion of this over the years. It has to do with classloading. The annotation processor classes need to be loaded into the compiler VM (or alternately, the entire Java model needs to be shipped over a VM boundary). For Eclipse, that means loading them into the IDE. Which means that once an annotation processor class is loaded, it effectively cannot be unloaded without restarting Eclipse.
The assumption has been that if an annotation processor is in the *project* path, it may well be something that the customer expects to change and recompile. That will not work, if we can't unload it after the first time it runs.
IntelliJ uses the javac compiler, launched in a separate process which terminates at the end of compilation. Thus it reloads every time (but, it does not get the benefit of incremental compilation or annotation processing).
Whether those assumptions are still true or relevant I can't say, but that's the background.
Perhaps more to the point, though, I do not think there is much in the way of active investment in the APT project. It was driven initially by BEA, who needed to support annotations for their Weblogic product and wanted to use the Eclipse IDE rather than their proprietary IDE. BEA contributed several developers for several years on this project (I was one of them). BEA no longer exists; I don't think these days there is anyone with such a strong vested interest in improving annotation processing; and the Java annotation processing API is cumbersome enough that I don't think many annotation processors have become mainstream.