[llvm] [AMDGPU] Re-implement splat scalar handling for packed 64-bit ops (PR #216896)

[Changpeng Fang via llvm-commits]

changpeng wrote:

Packed FP32 instructions may follow the same approach, and later optimization. But I would rather wait until the optimization is in place (We may see "NFC" effect then because we do not expect any test update).

https://github.com/llvm/llvm-project/pull/216896