Polygon model support

As part of milestone 2.0, we want to support polygon prediction and modeling. Here is a roadmap

• Add torchvision mask-rcnn to models/ directory and match the prediction format of boxes.
• Create a polygon dataset class.
  One area that I started implement in million trees is how the dataloader is going to yield masks. Frankly, I've never used a mask-rcnn model, https://github.com/weecology/MillionTrees/blob/64f78b52ad164fc4cbc87de2bf54259e98c694d4/milliontrees/datasets/TreePolygons.py#L134, so i'm open to ideas.
  (https://github.com/weecology/DeepForest/blob/d834b17b949e25df0f75905108955ec4928f025e/deepforest/dataset.py#L203)
• Confirm that preprocess.split_raster and other utilities/visualization functions can take in polygon predictions.
• Create a example ipython notebook showing how to train a polygon model, train an initial model using open source polygon data: https://www.mdpi.com/2072-4292/16/11/1935, https://zenodo.org/records/8148479 https://zenodo.org/records/8136161 (this one I had to clean up, the .tif was messy).

I expect that the user will not have to declare polygons anywhere, when they choose mask-rcnn in the config, or when we read_file we will know that they have polygon data. We may need to add a class attribute to deepforest.main that keeps track of the current geometry.

Optional ideas

• Compare output of mask RCNN with segment anything based on the DeepForest box model. see #460

Start by reviewing the models/ directory, the dataset implementation in deepforest/dataset.py, and the MillionTrees TreePolygons.py example to understand the proposed mask output. Then inspect preprocess.split_raster and the visualization utilities for polygon compatibility. Done would include mask-rcnn support, a polygon dataset class, compatible utilities, and an example notebook using polygon data.