Division by zero when training with small datasets caused by incomplete batch dropping

Nobody has claimed this yet.

I am aware that this repo is no longer maintained but I recently had to blow the dust off of it to replicate some stuff I did a few years ago. When setting up the environment I kept getting a division by 0 [here](https://github.com/mapbox/robosat/blob/cbb1c73328183afd2d6351b7bfa3f430b73103ea/robosat/tools/train.py#L197). I finally resolved it and thought I should leave a note here in case anyone finds themself in a similar situation.

It turns out it was because I was using a trivially small datasetr (only a few features) because I was just trying to get the entire pipeline working correctly. Specifically, when initializing the DataLoaders, the `drop_last` parameter is hardcoded to `True`, which will always drop the last batch if it is incomplete. So if you only have one (tiny) batch, you end up with an empty data loader, causing the aforementioned division by zero error.

Change this parameter to `False` [here](https://github.com/mapbox/robosat/blob/master/robosat/tools/train.py#L273) and [here](https://github.com/mapbox/robosat/blob/master/robosat/tools/train.py#L274) to fix it.

Start in robosat/tools/train.py, especially the division at line 197 and the DataLoader initialization around lines 273-274. Reproduce training with a dataset smaller than one batch, then verify that incomplete final batches are retained and training no longer fails with division by zero.

Written by the indexing model from the issue text.