Fixing trD compute type in the Xe Epilogue

[joyalbin]

This PR is to fix the trD_compute_frag type in Xe Epilogue based on the type configured in cst_callbacks.

ElementOutput in the LinearCombination and ElementOutput in the CollectiveEpilogue are two different things.
ElementOutput in CollectiveEpilogue is always same as ElementD, but the ElementOutput in LinearCombination need not to be same as ElementD.

ElementOutput in the LinearCombination represent the output of the LinearCombination operation (alphaAB, betaC).
At the same time, ElementOutput/ElementD in the CollectiveEpilogue represented the actual output which must be stored back.

[tdeng5]

Please add some UTs to cover enough scenarios, like: D=AxB + C => {FP32, FP16} = {FP32, FP16} + {FP32, FP16}

Please also combine #563 into this PR.