[AMD][DirectX ] WavePrefixProduct.* tests fail on AMD + DirectX

Nobody has claimed this yet.

WavePrefixProduct tests fail on AMD for floating point vectors of size 4 that are not 32 bits, and in the constant folding case.
A link to the logs:
https://github.com/llvm/offload-test-suite/actions/runs/24374208768/job/71184136405?pr=1074#step:13:688

A failure log snippet below:
```

# .---command stdout------------
# | Using DirectX API
# | Configuring execution on device: AMD Radeon RX 9070
# | RootSignature created.
# | Descriptor heap created.
# | Command structures created.
# | Creating SRV: { Size = 8, Register = t0, Space = 0 }
# | Creating UAV: { Size = 32, Register = u1, Space = 0, HasCounter = 0 }
# | Creating UAV: { Size = 32, Register = u2, Space = 0, HasCounter = 0 }
# | Creating UAV: { Size = 32, Register = u3, Space = 0, HasCounter = 0 }
# | Creating UAV: { Size = 32, Register = u4, Space = 0, HasCounter = 0 }
# | Creating UAV: { Size = 32, Register = u5, Space = 0, HasCounter = 0 }
# | EltSize = 8 NumElts = 1
# | SRV: HeapIdx = 0 EltSize = 8 NumElts = 1
# | EltSize = 8 NumElts = 4
# | UAV: HeapIdx = 1 EltSize = 8 NumElts = 4 HasCounter = 0
# | EltSize = 8 NumElts = 4
# | UAV: HeapIdx = 2 EltSize = 8 NumElts = 4 HasCounter = 0
# | EltSize = 8 NumElts = 4
# | UAV: HeapIdx = 3 EltSize = 8 NumElts = 4 HasCounter = 0
# | EltSize = 8 NumElts = 4
# | UAV: HeapIdx = 4 EltSize = 8 NumElts = 4 HasCounter = 0
# | EltSize = 8 NumElts = 4
# | UAV: HeapIdx = 5 EltSize = 8 NumElts = 4 HasCounter = 0
# | Buffers created.
# | PSO created.
# | Compute command list created.
# | Compute commands executed.
# | Read data back.
# `-----------------------------
# .---command stderr------------
# | Test failed: ExpectedOut4
# | Comparison Rule: BufferFloatULP
# | ULP: 2
# | Expected:
# | ---
# | Name: ExpectedOut4
# | Format: Float16
# | Stride: 8
# | Data: [ 0x3C00, 0x3C00, 0x3C00, 0x3C00, 0x4000, 0x4200, 0x4500,
# | 0x4700, 0x4400, 0x4880, 0x4E40, 0x5220, 0x4800, 0x4EC0,
# | 0x57D0, 0x5D5C ]
# | OutputProps:
# | Height: 0
# | Width: 0
# | Depth: 0
# | ...
# | Got:
# | ---
# | Name: Out4
# | Format: Float16
# | Stride: 8
# | Data: [ 0x3C00, 0x3C00, 0x3C00, 0x3C00, 0x4000, 0x4200, 0x4500,
# | 0x4700, 0x4400, 0x4880, 0x4E40, 0x5220, 0x4800, 0x4EC0,
# | 0x57D0, 0x4000 ]
# | OutputProps:
# | Height: 0
# | Width: 0
# | Depth: 0
# | ...
# | Full Hex representation of Expected Buffer Values:
# | [ 0x3c00, 0x3c00, 0x3c00, 0x3c00, 0x4000, 0x4200, 0x4500, 0x4700, 0x4400, 0x4880, 0x4e40, 0x5220, 0x4800, 0x4ec0, 0x57d0, 0x5d5c ]
# | Full Hex representation of Actual Buffer Values:
# | [ 0x3c00, 0x3c00, 0x3c00, 0x3c00, 0x4000, 0x4200, 0x4500, 0x4700, 0x4400, 0x4880, 0x4e40, 0x5220, 0x4800, 0x4ec0, 0x57d0, 0x4000 ]
# |
# | Test failed: ExpectedOut5
# | Comparison Rule: BufferFloatULP
# | ULP: 2
# | Expected:
# | ---
# | Name: ExpectedOut5
# | Format: Float16
# | Stride: 8
# | Data: [ 0x3C00, 0x3C00, 0x3C00, 0x3C00, 0x4000, 0x4200, 0x4500,
# | 0x4700, 0x4400, 0x4880, 0x4E40, 0x5220, 0x4800, 0x4EC0,
# | 0x57D0, 0x5D5C ]
# | OutputProps:
# | Height: 0
# | Width: 0
# | Depth: 0
# | ...
# | Got:
# | ---
# | Name: Out5
# | Format: Float16
# | Stride: 8
# | Data: [ 0x4200, 0x4500, 0x4700, 0x3C00, 0x4200, 0x4500, 0x4700,
# | 0x4700, 0x4200, 0x4500, 0x4700, 0x5220, 0x4200, 0x4500,
# | 0x4700, 0x5D5C ]
# | OutputProps:
# | Height: 0
# | Width: 0
# | Depth: 0
# | ...
# | Full Hex representation of Expected Buffer Values:
# | [ 0x3c00, 0x3c00, 0x3c00, 0x3c00, 0x4000, 0x4200, 0x4500, 0x4700, 0x4400, 0x4880, 0x4e40, 0x5220, 0x4800, 0x4ec0, 0x57d0, 0x5d5c ]
# | Full Hex representation of Actual Buffer Values:
# | [ 0x4200, 0x4500, 0x4700, 0x3c00, 0x4200, 0x4500, 0x4700, 0x4700, 0x4200, 0x4500, 0x4700, 0x5220, 0x4200, 0x4500, 0x4700, 0x5d5c ]
# |
# `-----------------------------
# error: command failed with exit status: 1

--

********************
```

Start by locating the WavePrefixProduct.* test definitions and the DirectX execution path, then reproduce the AMD failures using the linked workflow logs as reference. Compare the floating-point vector-size-4 and constant-folding cases with passing cases; done means these cases produce the expected buffers on AMD/DirectX without regressing other WavePrefixProduct tests.

Written by the indexing model from the issue text.