[LT] Copy the device ordinal in MakeComputationDataFromTensor (#72574)
Summary:
Currently, even if we specify 'lazy:1' for tensors, the tensors will still
be running on the default ordinal, i.e., 'lazy:0'. To fix that, we need to
copy the device ordinal while constructing computation data.
Test Plan:
Manual test that uses nvidia-smi to monitor if the running program is
on the correct GPU.