[llvm] [AArch64] Expand FORM_TRANSPOSED_REG_TUPLE to copies before regalloc (PR #207205)

[Benjamin Maxwell via llvm-commits]

MacDue wrote:

> I think the problem isn't that the psuedo doesn't model enough (that's a choice), but rather that its expansion into COPY instructions is flawed and currently makes unsafe assumptions.

I don't think the pseudo allowing hard to expand cases was an intentional design choice (or desired), so I think that is a problem from the perspective of this PR.

https://github.com/llvm/llvm-project/pull/207205