Add random sampling scheme

[jejjohnson]

Taking inspiration from rastervision, our current implemention is the equivalent of a sliding window dataset. However, I see a very nice use case where we could use a random sampler to randomly sample patches. This would be useful for training whereby we do not need any criteria for sampling. This also makes training easier for cases where we can simply training with maximum iterations instead of epochs.

Possible Add-On - we could also sample according to some criteria (perhaps user specified). For example, a brightest pixel function or some other criteria.

TorchGeo - Sampler API - code