[llvm] [AMDGPU] Use larger immediate values in S_NOP (PR #158990)

[Jay Foad via llvm-commits]

jayfoad wrote:

> Probably should include a machine verifier check the correct number of bits are used

That would break `test/CodeGen/AMDGPU/llvm.amdgcn.s.nop.ll` which includes:
```
  call void @llvm.amdgcn.s.nop(i16 63)
```
I guess that should either be disallowed, or it should generate multiple s_nop instructions, but I didn't want to get into changing the behaviour of an intrinsic in this patch.

https://github.com/llvm/llvm-project/pull/158990