Fix reprojection issues

[adamjstewart]

This PR fixes 2 issues related to reprojection:

• RasterDataset: fix loading of files requiring reprojection

I broke this in #1329. When loading any file that requires changing the CRS (and therefore WarpedVRT), we see the following error:

Traceback (most recent call last):
  File "/Users/Adam/torchgeo/test.py", line 11, in <module>
    ds[ds.bounds]
  File "/Users/Adam/torchgeo/torchgeo/datasets/geo.py", line 899, in __getitem__
    samples = [ds[query] for ds in self.datasets]
  File "/Users/Adam/torchgeo/torchgeo/datasets/geo.py", line 899, in <listcomp>
    samples = [ds[query] for ds in self.datasets]
  File "/Users/Adam/torchgeo/torchgeo/datasets/geo.py", line 443, in __getitem__
    data = self._merge_files(filepaths, query, self.band_indexes)
  File "/Users/Adam/torchgeo/torchgeo/datasets/geo.py", line 486, in _merge_files
    dest = src.read(
  File "rasterio/_warp.pyx", line 1233, in rasterio._warp.WarpedVRTReaderBase.read
ValueError: WarpedVRT does not permit boundless reads

The fix is to always use rasterio.merge.merge instead of src.read, even if there is only a single file. This bug wasn't detected because we don't actually load real files and change the CRS during the RasterDataset tests. I added real files and ensured that we test loading those files with a different CRS.

• Intersection/UnionDataset: fix crs/res propagation

This bug was reported by @jatentaki in #1341. Basically, the following works fine:

ds = (ds1 & ds2) & ds3
ds = (ds1 | ds2) | ds3

However, the following is completely broken:

ds = ds1 & (ds2 & ds3)
ds = ds1 | (ds2 | ds3)

The fix for this was much more involved. We need to add a res getter/setter to GeoDataset, similar to our crs getter/setter. Then we need to override this in Intersection/UnionDataset to update both datasets. This is updated recursively to ensure that all datasets have matching crs/res. This is a pretty uncommon use case (ds1 & ds2 & ds3 maps to the former), so that's why no one noticed until now. I added tests for both of the above situations so this doesn't break again.

Thanks @jatentaki for reporting this!

Fixes #1341

[adamjstewart]

Although the getter/setter were added in 0.2, the attribute itself has always been present, so technically there was no API change as far as the user is concerned.

[calebrob6]

My only concern here is on performance but let's sort that out later. I'm imagining that we can have something like benchmark.py run during integration tests to test for dataloader performance regressions.

[adamjstewart]

I recall it being slightly slower (that's why we added the if-statement in the first place), but I'm unaware of any other way of doing this with src.read, so for now if we want boundless reading (needed for GridGeoSampler) then we have to use rasterio.merge.merge.

[calebrob6]

Sure, but it still stands that this is a place that needs revisiting

[calebrob6]

Sure, but it still stands that this is a place that needs revisiting