Peak positions can be out-of-frame

[matbryan52]

While testing #70 we've noticed that in certain situations evaluate_correlations can give peak coordinates which are outside of the frame if the input peaks are close to the frame edges and the crop_size therefore overlaps the border. This happens because we simply _shift the coordinate in the crop to the coordinate in the frame, without applying any bounds checking.

This doesn't have much impact on internal code as we refine/evaluate in the coordinate system of the crops, but could affect downstream code which might assume results will lie in the frame. I'd argue this is unexpected because the cropping / refinement process is quite opaque to the user (it's mostly just the search parameter on a MatchPattern and that isn't really clear in itself).

Another potential error (which I haven't actually observed!) is the fact we create the centers buffer as np.uint16, so this could overflow in some situations.

One option might be to supply float32 / np.nan results even for the integer centers to make this error case easier to identify.