fix(hnsw): handle invalid keys in group-by query with filter (#781)

[thlurte]

Problem

Fixes #781.

When executing Collection.group_by_query with a scalar filter predicate on an HNSW index, the query returned empty results with error logs:

ID is out of range: id[18446744073709551615] count[...]
Fetch group by field failed: field[...] row_id[18446744073709551615]

Root Cause:

1. During HNSW group neighbor expansion (expand_neighbors_by_group), search heap nodes may return kInvalidKey (UINT64_MAX) when unmapped or deleted.
2. Passing UINT64_MAX to DocFilter::is_filtered() caused an out-of-bounds access in the bitmap and returned nullopt (unfiltered).
3. The unfiltered invalid key was then passed to VectorRecallNode::group_fun, which invoked segment_->fetch with row_id = UINT64_MAX (-1), failing to fetch the group field and aborting group formation.

Solution

1. HNSW Algorithms (hnsw_algorithm.cc, hnsw_sparse_algorithm.cc, hnsw_rabitq_query_algorithm.cc):
   • Skip kInvalidKey during initial group partitioning.
   • Guard candidate queue refill against kInvalidKey.
   • Update filter lambda to treat kInvalidKey as filtered (true), preventing unmapped nodes from being selected or expanded.
2. DocFilter (doc_filter.cc):
   • Add boundary check if (id == std::numeric_limits<uint64_t>::max() || id >= segment_->doc_count()) return true; to prevent bitmap index overruns.
3. VectorRecallNode (vector_recall_node.cc):
   • Guard group_fun against invalid row_id / UINT64_MAX before calling segment_->fetch().
4. Tests (collection_test.cc):
   • Added regression test Feature_GroupByQuery_HnswWithFilter verifying filtered group_by_query on HNSW index for both in-memory and mmap collections.