Fixes T11743. For very large inputs which can't be simplified (e.g., dissimilar text at the beginning and end across a very large number of paragraphs) we currently try to build an edit distance matrix, but this is exponential in space and time and slowness and how PHP-ish it is.
Instead, just give up for very large inputs. This will still prose-diff any corpuses with fewer than 128 paragraphs, which is the vast majority of documents.