[llvm] [AMDGPU] NFC, fix geps, add vector-combine pass to promote-alloca-array-to-vector test (PR #166280)

[via llvm-commits]

choikwa wrote:

SWDEV-559837 has shown that promote-alloca generates legalization code that uses element by element copy. My intention is to fix this in the Vector Combiner, combining chain of extract-insert vectors into few shufflevectors. foldInsExtVectorToShuffle already exists but it only looks at a pair of extract-insert, and cost is not favorable for all but first element copy.

https://github.com/llvm/llvm-project/pull/166280