Canonicalization of GEPs to i8

LLVM now canonicalizes some (constant) GEPs to perform address calculations in terms of bytes. This is part of larger change of moving towards `ptradd` instead of `getelementptr`. This will require changes in clspv to keep pace.

At the moment I'm seeing different failures:
* unit tests with incorrect codegen expectations (~115 per arch)
* cts failures that look functionally incorrect

For example, in `test/CommonBuiltins/min/half2_fmin.cl` the loads are broken down into 4 1-byte loads, a vector creation, and a bitcast from `v4uchar` to `v2half`. This is problematic for a few reasons:
1. Not all devices support byte addressing and clspv now requires it.
2. It relies on downstream drivers to properly coalesce memory accesses. This might be ok, but I don't have performance numbers that are easily accessible one way or the other.
3. It muddies the IR. This isn't *just* an aesthetic concern. There is a binary size increase in these cases due to more instructions getting used.

CTS failures:
* basic/vstore_local (fails char2 and uchar2)
* relationals/shuffle_copy
* relationals/shuffle_function_call
* relationals/shuffle_built_in
* relationals/shuffle_built_in_dual_input
* relationals/shuffle_array_cast

The likely solution is a change in type inference that can find a better type than i8 by looking at the GEP's users (right now it stops at any GEP).