[AMDGPU] Fix LOD bias in A16 combine

As the codegen fix in D111754, the LOD bias needs to be converted to 16
bits. Fix this in the combine.

Differential Revision: https://reviews.llvm.org/D116038