The difference should be that the RealICE emulates the microcontroller, so there is no physical presence of the target mcu needed in the circuit.
At first i would assume that for in circuit debugging without RealICE, you need separate space in the program memory. Secondly i assume from the prior statment, that for debugging with ICD or PicKit something has to be computed in the PIC in the background to make debugging possible.
So RealICE should be way faster.
If the assumption that the debugging program reserves space in the MCUs memory is true, then you will never be able to debug smaller PICs with any ICD or PicKit tool.
Edit: it seems to be somewhat different to conventional emulators like ICE2000. As i remember, ICE2000 had several target boards, specific for the desired device. RealICE seems to work differetly...
The Real Ice isn't an emulator as you suspected. I've got one and an ICD3 and honestly don't make any use of the extras that the Real Ice offers. There are a few cool things like a hardware trace real time watch and I/O capability but they aren't well developed or supported.
You are right in that many of the smaller processors require some reserved space for debugging but most of their processors now have dedicated internal hardware which keeps the debugging out of the way. Certainly there are less overall resources at the low end of the processors.
I honestly don't know why the call it an emulator since it requires a processor to do anything. What does it "emulate?"