[Feature] Enable CuPy on main process

• Make GPU processing available for operations on the main node (chiefly libertem.udf.base.UDF.merge() and libertem.udf.base.UDF.get_results()) through libertem.udf.base.UDF.xp. This is targeted at methods like iDPC and iCoM where the result is calculated in libertem.udf.base.UDF.get_results() with Fourier transforms over the displacement vector field that run much faster on GPUs compared to CPUs. It can be controlled with the new main_process_gpu parameter for libertem.api.Context.make_with(). It is enabled by default to catch potential issues, to be reviewed before the next release (#1759).