Add support for packed registers with vectors.
Example:
```
%wo0 = nvvm.inline_ptx
"dp4a.s32.s32 {$w0}, {$r0}, {$r1}, {$r2};"
ro(%src, %mask, %zero : vector<4xi8>, i32, i32)
-> i32
```
Here, `vector<4xi8>` is lowered to an `i32` register (i.e., an `r` in
PTX).