Remove upcasting workaround for FFT operations once oneMKL introduce native support for Half dtype.

## Problem description

There was a problem that MKL FFT doesn't support dtype Half. It was originally tracked in the issue #2617.
There was a workaround added that simply upcasts the Half dtype to single precision 32 bit type and then performs calculations. This workaround was introduced in Pull Request https://github.com/intel/torch-xpu-ops/pull/2637. That is mathematically valid and correct, but for the performance reasons it would be better to use directly the native MKL Half dtype computations for FFT.

## Task

The Half dtype support for MKL FFT kernels is on the MKL roadmap.
This task issue is to track the MKL releases and remove the workaround introduced by #2637 once MKL oficially supports the Half dtype for FFT operations.