…nse#2806)

Deep pagination with --enable-lazy-filter returns empty hits because
the topster buffer is sized using approx_filter_ids_length, which
undercounts for NOT-IN array filters. The OR-sum of excluded posting
lists overcounts overlapping docs, causing num_ids - OR_sum to
underestimate the NOT result size.

String path: use num_ids - max(posting_list_sizes) instead of
num_ids - OR_sum. This is provably safe (union >= max(Si)) and
preserves exact behavior for single-value NOT-equals. The OR-sum
hint is preserved for the eager-materialization threshold check.

Int/Float paths: set approx to num_ids (the accumulation structure
differs — each list already stores num_ids - Si, not raw Si). The
OR-sum hint is separated for threshold decisions.

Add regression test: 100 docs with compound NOT-IN filter, verifying
all 70 matching results are retrievable across all pages.

Multi-token string NOT filters can undercount lazy pagination because token-level estimates are not safe until exact value matching runs. Use a conservative bound for lazy NOT iterators and add regression coverage for deep pagination.

Add LazyFilterNotInNumericOverlapDeepPagination test that exercises the
lazy NOT-filter approximation bug with overlapping int32[] posting lists.

10 docs with tags=[1,2] create overlapping posting lists (S1=S2=10).
OR-sum=20 causes old approx = num_ids-20 = 5 (underestimate vs actual 15).
With approx<=20 in TEST_BUILD, compute_iterators is skipped, topster is
capped at 5, and page 2 returns 0 hits instead of 5.

Verified: fails on v31 without fix, passes with PR typesense#2812 fix applied.