[analyzer] Revert #115918, so empty base class optimization works again (#157480)

Tldr;
We can't unconditionally trivially copy empty classes because that would
clobber the stored entries in the object that the optimized empty class
overlaps with.

This regression was introduced by #115918, which introduced other
clobbering issues, like the handling of `[[no_unique_address]]` fields
in #137252.

Read issue #157467 for the detailed explanation, but in short, I'd
propose reverting the original patch because these was a lot of problems
with it for arguably not much gain.
In particular, that patch was motivated by unifying the handling of
classes so that copy events would be triggered for a class no matter if
it had data members or not.
So in hindsight, it was not worth it.

I plan to backport this to clang-21 as well, and mention in the release
notes that this should fix the regression from clang-20.

PS: Also an interesting read [D43714](https://reviews.llvm.org/D43714)
in hindsight.

Fixes #157467
CPP-6574