Fix IndexIVFFlatDedup::remove_ids not counting duplicate-id removals

[trang-nm-nguyen]

Summary:
https://www.internalfb.com/monitoring/oncall_operator?board=1003685189388103&workItem=1079527385016771

---

IndexIVFFlatDedup::add_with_ids increments ntotal for every added vector,
including duplicates (n_add++ runs for every id, then ntotal += n_add). A
duplicate vector is not stored as a new inverted-list entry; it is recorded
only as an entry in the instances multimap that maps a stored representative
id to the duplicate id.

remove_ids, however, derived its return value and its ntotal decrement
solely from nremove — the count of physically shrunk inverted-list entries.
Removing a duplicate-only id (an id that appears only as a value in instances)
erases the instances entry but shrinks no list, so nremove stayed 0:
remove_ids returned 0 and left ntotal unchanged even though a valid id was
removed. The same undercount happened when a stored representative id was
removed but a surviving duplicate took over its slot (the slot is rewritten,
not shrunk).

The library maintains the invariant
ntotal == (physical inverted-list entries) + instances.size(), so the true
number of logically removed ids is
nremove + (instances_before - instances_after). This captures both the
physically removed entries and the net drop in instances (erasures minus the
toadd re-links used when a representative with several duplicates is removed).
The fix records instances.size() before the rebalance and adjusts the count
and ntotal decrement accordingly.

Differential Revision: D118606761

[meta-codesync]

@trang-nm-nguyen has exported this pull request. If you are a Meta employee, you can view the originating Diff in D118606761.