[clang] [ClangFE] Improve handling of casting of atomic memory operations. (PR #86691)

[Matt Arsenault via cfe-commits]

arsenm wrote:

> I think this case isn't that simple as it is an 80 bit value. Currently that is loaded atomically first with i128, then stored as a temporary and then loaded as an fp80. If I remove that casting, the verifier complains "atomic memory access' operand must have a power-of-two size".

I think this should be more targeted based on the bitesize. PPCf128 should be simple 

https://github.com/llvm/llvm-project/pull/86691