Post processing of predicted labels using patch context

[rwood-97]

As per #287 , instead of using patch context during training/inference stage it would be good to add a post-processing step which sanity checks the predicted labels.

Consider examples (X = railspace, o = no):

o o o
o X o
o o o
^ This is likely a false positive

o o o
X X X
o o o
^ This is likely correct

o o o
o X o
o X X
^ This likely requires attention/review

This works particularly well for linear features such as railspace.
See hackmd notes and #287 for more info on this.

Describe the solution you'd like
First steps would be to create this as part of a script and have a go with 2nd ed patches and labels.
We should then look through results and make any adjustments as needed.

Second steps would be to integrate this into the MapReader code itself.
The main thing here is to think about the broader use cases of MapReader and how to make sure any code that is inbuilt is generalisable.