Modify the CShuffleEpilogue test to initialize input values on the host
side and load them into the device's acc_tile register block, rather
than generating values directly on the device.
Changes:
- Add input_data parameter to test_cshuffle_epilogue_kernel
- Create input tensor view and tile window with epilogue distribution
- Load input from global memory into acc_tile using load_tile()
- Update launch helper functions to accept and pass device_input
- Update run_cshuffle_epilogue_test to create host input tensor,
copy to device, and return both input and output
- Update run_scale_comparison_test to handle new return type
Co-Authored-By: Claude <noreply@anthropic.com>
670fb88e82