As reported in GH61959, the patch for https://reviews.llvm.org/D146178
regressed some comparisons of non-out-of-line function constraints.
This patch fixes it by making sure we skip the list of template
arguments from a partial specialization (where they don't really
matter!).
There was 1 other workaround that was made for checking deduced
arguments that this removes part of.
Fixes: #61959
It's not obvious to me what this was doing, so I'm not sure whether it's correct to remove it. Can you explain? It makes me uncomfortable that we would treat class template partial specializations and variable template partial specializations differently, at least without a good explanation for why that's appropriate -- perhaps we should apply this same set of changes to variable template partial specializations too, and remove this mechanism entirely?