Geosampler prechipping

[sfalkena]

I think it makes sense to be able to interact with the chips of each sampler. Therefore, I want to propose changing the creation of the chips in the __init__ of each sampler, and to keep track of the chips using a GeoDataFrame. This leads to the following benefits:

• It is possible to save the chips to file and to visualize them either in GIS software, or while developing.
• It allows to filter the chips using other geometry, to remove chips that are not of interest. This could be particularly useful during inference to limit the number of chips. Example: When I want to infer a building detection model, I can now instantiate a gridsampler, filter my chips using a shapefile of urban areas and only infer on the intersection of my chips with the urban areas. See the function filter_chips.
• It allows to split the samples. Right now I have created a function to split the samples to spread them across multiple workers. See the function set_worker_split.

An abstract method needs to be implemented for every sampler:

    @abc.abstractmethod    
    def get_chips(self) -> GeoDataFrame:

[sfalkena]

See mirrored PR to TorchGeo: microsoft#2300