[DXIL] Broken codegen for loading elements from FP16 matrix types in StructuredBuffer

On latest main I get the wrong result when loading elements from fp16 matrix types stored in structured buffers.

struct Data
{
    float16_t4x4 a;
};

StructuredBuffer<Data> buf : register(t1);

...
float16_t2 b = buf.a[0].xy;

This loads elements a[0][0] and a[0][2], not a[0][1] as expected.
Looking at the generated code it looks like it thinks there is 4B between each element not 2B.

See attached example HLSL and generated DXIL.
dxc-float16_t4x4-example.zip

Start with the attached dxc-float16_t4x4-example.zip and inspect the generated DXIL for the StructuredBuffer access in the provided HLSL. Reproduce the load of buf.a[0].xy and trace the element offset calculation; done means the generated code loads adjacent FP16 elements at 2-byte spacing and produces the expected values.