Filter duplicate rows when pruning vectors

[afonsojanu]

Fixes#5397

Vocab.prune_vectors builds its keep/discard split from one priority entry per lexeme, but more than one key can point at the same vector row. That happens for example after Vectors.add(key, row=...) aliases a spelling or casing variant onto an existing row instead of duplicating the data, which is a common way to save memory in real vector tables.

When two keys share a row, the current code can end up copying that row twice into the pruned table (wasting a slot that could have held a genuinely distinct vector), or splitting the pair across the kept and discarded halves so one of them gets sent through the nearest-neighbour search below even though its exact vector is still sitting right there among the rows that were kept.

This collapses the priority list down to one entry per row before deciding what to keep, then afterwards restores every aliased key by pointing it at wherever its row's representative key ended up, whether that representative was kept in place or itself remapped to a synonym.

Added test_vocab_prune_vectors_duplicate_rows alongside the existing test_vocab_prune_vectors: it builds a vocab where kitty/puppy alias the same rows as cat/dog, and checks that pruning to 2 rows keeps both pairs' vectors intact (rather than wasting a slot duplicating one of them) while the one truly distinct vector still gets remapped to its nearest neighbour as before.

I reverted just the vocab.pyx change locally and confirmed the new test fails with the exact symptom described in the issue (dog/puppy's shared row gets split, with puppy landing in the remap output alongside the genuinely discarded word) before restoring the fix.

Ran the full spacy/tests/vocab_vectors/ suite (106 passed) and cython-lint on vocab.pyx (same 19 pre-existing warnings as upstream master, no new ones introduced).