pass subvolume slicing to TemplateMatchingGPU for improved std

[McHaillet]

Currently for tomograms that are tiled into patches for GPU memory constraints are used fully to calculate the standard deviation of the search. However, there always needs to be overhang between patches (namely template_size // 2) to get an accurate score map for the whole tomo. This results in standard deviations being averaged over areas that are calculated multiple times, leading to a incorrect final search standard deviation.

It could be easily fixed by passing a slice to the template matching run based on indexes that are already present to place the tile back into the full tomo, and only calculate the std in the sliced area.