May be I understand all that wrong, but that is what I think:
a) mastering plug-ins (with huge latency)
do not depend from / do not influence
b) hardware latency / buffer size
at least in theory (when (a) are written correctly)
Mastering plug-ins delay output by significant time, >>0.01sec. Free VST announced in another thread can show exact delay (Sonar does not expose that number, while it is aware about it). Such plug-ins are not "slow" because they need more CPU time, they just need more information "from the future" to make the decision. They can be CPU intensive or consume close to no CPU at all, as any other plug-ins. Some synths which are supposed to work life, can be rather CPU hungry. All plug-ins require some minimum system power to avoid glitches (which IS buffer size dependent), but absolute amount of that power is not plug-in type dependent. Important that after the level is reached, "normal" plug-ins produce 0 latency (the output buffer correspond to the current input buffer) while mastering plug-ins produce constant big latency (the output buffer correspond to the input buffer sent ~0.7 second ago for CW LP with default settings).
The conclusion from that part:
with LP Comp. inserted, when you change something, you hear the result after 0.7 of a second. By "something" I mean not only life MIDI input but also volume changes, synth/FX parameter changes, etc.
That is normal, by design, by definition of what these plug-ins do, nothing can be "fixed" nor speedup there.
About CPU consumption, the buffer size and mastering plug-ins. Some "simple" plug-ins which work just with one buffer theoretically can consume less CPU power when the buffer size is decreased, so if let say you have changed the size from 256 to 128, such plug-in takes 2 times less CPU power per buffer and so overall CPU consumption (per second) stay the same. In practice DAW/Hardware/Drivers add significant overhead and most plug-ins are not so simple, so we observe some CPU increase. Mastering plug-ins are working with more then one buffer per time, so when we decrees the buffer size from 256 to 128, the amount of audio information and so CPU such plug-in need stay the same. But it will be called 2 times more per second, so
for mastering plug-ins 2 times more CPU consumption can be expected when reducing the buffer by half.
But please note that "big" buffer size you set in Sonar has nothing to do with the size such plug-ins need to process each time, for example 2048 buffer size for 44.1kHz is just 0,05 of a second while LP Comp pretend to use 0,7. Also the buffer "delay" should be added to the effect delay to get expected delay, in the example we can expect 0,75 sec.
With all that written, current version of CW LP plug-ins have a bug in the GUI design. Also that GUI can have some compatibility issues with graphic hardware, which can trigger audio clicks/glitches. That has nothing to do with your CPU power. And so,
when reporting problems with new plug-ins, please check either these problems still exist when there GUI is closed.
I repeat, I do not pretend I am right with all that. That is my personal opinion only (but it is based on practical experience with VST/DX plug-ins development plus some signal processing background).