Currently, the ptrmask intrinsic allows the mask to have any
size and will zero-extend or truncate it to the pointer size.
However, per the specified semantics, what we would actually
need to do is to first zero-extend or truncate it to the pointer
index size and then 1-extend it to the pointer size. This seems
to leave a lot of room for error, so this patch proposes to
make the intrinsic stricter:

It now requires that the mask type matches the pointer index type
-- a zext or trunc can be done explicitly in IR and should not
be part of the intrinsic. Also spell out that the mask is 1-extended
to the pointer size if we're talking about the integer representation
(this is implied by the GEP expansion).