[DirectX] Add with overflow should emulate for everything except specifically 32-bit ops

Our current add with overflow implementation always tries to use the 32-bit dxil ops but this only actually makes sense when doing 32 bit adds. Currently this means 16 and 64 bit operations are incorrect.

Locate the DirectX add-with-overflow implementation and the entry point where it selects 32-bit DXIL operations. Compare the behavior for 16-, 32-, and 64-bit adds; done means non-32-bit operations emulate overflow correctly while 32-bit operations retain the DXIL path.

Written by the indexing model from the issue text.