Hi and sorry about this – it's currently expected that the interface becomes less performant for very long documents with many tokens, and we're working on a rewrite that doesn't have this problem. I think what makes it additionally tricky in your case is that you just end up with more tokens overall, due to the way the characters are segmented.
As a workaround, one thing you can do is use patterns to disable all tokens that you know won't ever be part of a relation – of course, only if that's possible. Obvious candidates are punctuation, but you might be able to write other disable patterns based on part-of-speech tags etc.
Can you double-check that when you load your custom pipeline with the tokenizer in Python and process a text, the tokens are segmented correctly? If a Doc produced by the model shows the correct tokens, Prodigy should refelct this accordingly in all recipes that use the model for tokenization.