Create pipeline for inference with gridded data to create map plots of ML prediction

[stevehadd]

Currently we're focused on training and inference with the tabular dataset. Ultimately we want to display predictions as maps. We thus want to be able to load some gridded data (as would come directly out of the model and stage pipeline), and be able to easily do inference to create a ML prediction. This pipeline doesn't exiust yet so should be created in anticipation of further sharing of our results.

[stevehadd]

it would be good to create a dataset class for gridded data, where like with tabular data you can get a memory object (possibly lazily loaded) back into memory. This could be backed by zarr and return an xarray object like one gets a pandas dataframe.

Would also be good to include more of lazy loading paradigm into the datasets concept (like maybe create a dask or ray or DVC cluster through the AzML compute GUI) to do you processing and then it returns your reduced in memory object.

Would be good to have different specialised datasets e.g. geospatial data, satellite images, 3D point cloud etc.