Create tests to assert coherence between inference and evaluation mode.

[bw4sz]

Given the current changes to the datasets class and questions about resizing #1240 I think we should have tests that assert that main.predict_file and trainer.validate produce the same predictions. As well as a a test that asserts that preprocess.split_raster + trainer.validate has the same behavior as main.predict_tile. There a couple caveats to the tiling question because the windows aren't the same for other dataloader strategies (

DeepForest/src/deepforest/main.py

Line 516 in e44b268

dataloader_strategy="single",

) but it should be true for the default 'single' window.