AMDGPU/GlobalISel: Fix get.rounding s_getreg lowering (#205601)
Use llvm.amdgcn.s.getreg instead of emitting S_GETREG_B32 directly so
instruction selection applies the required SReg_32 operand constraint.
This was done for setreg but missed for getreg.
Fixes https://github.com/llvm/llvm-project/pull/205265 when expensive
checks are enabled.