fix: stemming

[stephantul]

The stemming algorithm in the original model was applied after the original terms were indexed. This resulted in indexing errors if two different terms had the same stem.

Before:

_, y = d.process_document("hello hellos")
print(y)
# {"hello": 2}

Now:

_, y = d.process_document("hello hellos")
print(y)
# {"hello": 1}

This raises scores on Nanobeir a tiny bit.