Implement `PVDataSource.get_locations`

[JackKelly]

Detailed Description

Each example should include at least one PV system. These PV systems don't have to be in the center of the satellite imagery. This way, we get a huge number of potential central positions for each example (far larger than the number of PV systems).

Context

More variety is almost certainly necessary to reduce over-fitting.

Possible Implementation

For each example:

1. Pick a random PV system
2. Move the center of the region of interest by some random amount, such that the PV system is still guaranteed to appear within the ROI, inside a border. e.g.: x += rng.uniform(low=-wiggle_room, high=wiggle_room) where wiggle_room = (width_of_satellite_region_of_interest / 2) * 0.8 (where 0.8 is just a guess... we don't want the PV system to be right on the very edge of the region of interest).
3. Do the same for y.

This suggests that PVDataSource needs to know the size of the satellite region of interest.