[AMDGPU] Use vector memory types for 64/128-bit cooperative atomics (#229501)
The 16x8B and 8x16B cooperative atomic intrinsics had integer memory
types (`i64`/`i128`) that didn't match their vector values, which broke
value tracking on the loaded value and could crash the compiler. This
patch uses the value type as the memory type and updates the selection
patterns to match. Codegen for existing cooperative atomics is
unchanged.