Properly implement DynamicMap.split_overlays

[philippjfr]

As I originally described in #2253, the DynamicMap.split_overlays method simply reuses the HoloMap implementation of the method which is not correct. This is because it will simply split up the Overlays in the cache into their constituent components and then make a clone of the DynamicMap. This means that if anything attempts to actually call that DynamicMap (which generally shouldn't happen but might) the validation on the DynamicMap will raise an error because the type returned by the callback (an Overlay) will not match the type in the cache (an Element).

The correct solution is probably just to implement a split_overlays method that fetches the appropriate layer from the original DynamicMap using something like:

def split_overlays(self):
    keys = self.last.keys()
    dmaps = []
    for i, key in enumerate(keys):
        dmaps.append(self.map(lambda overlay: overlay.get(i), Overlay))
    return keys, dmaps