I'm not worried at all about 2 channel memory bandwidth.
If I were, I'd go with Kabylake-X.
I'm not understanding what you're getting at then.
You believe Coffee Lake will bring higher IPC because of RAM bandwidth.
Kaby Lake and Skylake are already pushing up against the limits of DDR4, so Coffee Lake won't be able to provide higher bandwidth.
Now you say you don't care about increasing RAM bandwidth? I don't understand.
EDIT: Forgot, but Kaby Lake-X has the same dual-channel controller as Kaby Lake does, so that's also an irrelevant point. The point I'm trying to get at is that Skylake and Kaby Lake, short of some miracle advance in DDR4 manufacturing, have already pushed dual-channel DDR4 to its practical limit. If you're concerned with getting more bandwidth (as you implied by saying Coffee Lake would allow higher RAM bandwidth), then the only real option is to move to a quad-channel platform.
Also, process refinement and die shrinks don't inherently require tweaking the way Pascal did - it was not a direct die shrink of Maxwell. An example of a direct die shrink is Prescott-Cedar Mill. Cedar Mill was only "better" due to lower power consumption, which was a direct result of the die shrink.
Intel's deliberate changes from Skylake-Kaby Lake-Coffee Lake are quite similar to what Intel would traditionally call CPU steppings - such as Cedar Mill's B1, C1, and D0. They were all process refinements, but Intel made no effort to market them as new products.