[AMDGPU] Split unaligned LDS access instead of scalarizing

There is no need to fully scalarize an unaligned operation in
some case, just split it to alignment.

Differential Revision: https://reviews.llvm.org/D123330